Congrats to @[ods.ai] topcoders, @jacobkie, @Deep Retina, @Nuclear Vision, @Inom Mirzaev for being the top 5 and all kagglers who had completed the challenge.
\n\nIt has been a long and tough challenge. All has done well! I think the scores of the top 5 are amazing and their models did generalise very well. Can't wait to hear more about their solutions!
\n\nSo UNet Vs Mask-rcnn, which is better?
", "messages": [ { "id": "315281", "postDate": "04/17/2018 00:41:15", "content": "Congrats to @[ods.ai] topcoders, @jacobkie, @Deep Retina, @Nuclear Vision, @Inom Mirzaev for being the top 5 and all kagglers who had completed the challenge.
\n\nIt has been a long and tough challenge. All has done well! I think the scores of the top 5 are amazing and their models did generalise very well. Can't wait to hear more about their solutions!
\n\nSo UNet Vs Mask-rcnn, which is better?
", "rawMarkdown": "Congrats to @[ods.ai] topcoders, @jacobkie, @Deep Retina, @Nuclear Vision, @Inom Mirzaev for being the top 5 and all kagglers who had completed the challenge.\n\nIt has been a long and tough challenge. All has done well! I think the scores of the top 5 are amazing and their models did generalise very well. Can't wait to hear more about their solutions!\n\nSo UNet Vs Mask-rcnn, which is better?", "votes": null }, { "id": "315285", "postDate": "04/17/2018 00:44:42", "content": "Congratulations to all winners! Well done!
\n\nSpecial thank you to @ Heng CherKeng for the excellent code, discussions and generous contributions. The competition would not be so interesting if it was not for these contributions.
", "rawMarkdown": "Congratulations to all winners! Well done!\n\nSpecial thank you to @ Heng CherKeng for the excellent code, discussions and generous contributions. The competition would not be so interesting if it was not for these contributions.", "votes": null }, { "id": "315286", "postDate": "04/17/2018 00:47:00", "content": "\n\n\nSo UNet Vs Mask-rcnn, which is better?
\n
Maybe a combination of them?
\n\nFrom my computation based on stage1 participation, you're one short of gold medal, which is a pity given all you shared. I can't say I hope someone will be removed, but you really deserve a gold medal here.
", "rawMarkdown": "> So UNet Vs Mask-rcnn, which is better?\n\nMaybe a combination of them?\n\nFrom my computation based on stage1 participation, you're one short of gold medal, which is a pity given all you shared. I can't say I hope someone will be removed, but you really deserve a gold medal here.", "votes": null }, { "id": "315288", "postDate": "04/17/2018 00:48:35", "content": "@CPMP no worry. there is always the next kaggle challenge!
", "rawMarkdown": "CPMP no worry. there is always the next kaggle challenge!", "votes": null }, { "id": "315292", "postDate": "04/17/2018 00:51:48", "content": "Has anyone tried something other than U-Net or Mask-RCNN and got a score above 0.500?
", "rawMarkdown": "Has anyone tried something other than U-Net or Mask-RCNN and got a score above 0.500?", "votes": null }, { "id": "315294", "postDate": "04/17/2018 00:52:24", "content": "Thanks again Heng for really embodying the true spirit of Kaggle. My current impression is modifications of UNet are a little better for this problem--that's what we used. Allen just reported 0.582 private for his best Mask-RCNN model; anybody beat that? @CPMP I was dying to try and find a Mask-RCNN result to ensemble with our good UNet but never found one that passed comparable muster locally.
", "rawMarkdown": "Thanks again Heng for really embodying the true spirit of Kaggle. My current impression is modifications of UNet are a little better for this problem--that's what we used. Allen just reported 0.582 private for his best Mask-RCNN model; anybody beat that? @CPMP I was *dying* to try and find a Mask-RCNN result to ensemble with our good UNet but never found one that passed comparable muster locally.", "votes": null }, { "id": "315308", "postDate": "04/17/2018 01:04:54", "content": "we've managed to get 0.585 with a single Mask-RCNN, but with late submission :(
\n\n", "rawMarkdown": "we've managed to get 0.585 with a single Mask-RCNN, but with late submission :( \n\nhttp://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png", "votes": null }, { "id": "315325", "postDate": "04/17/2018 01:19:07", "content": "@Russ W, Thanks for sharing, can't wait for a more complete write up form your team!
", "rawMarkdown": "Russ W, Thanks for sharing, can't wait for a more complete write up form your team!", "votes": null }, { "id": "315326", "postDate": "04/17/2018 01:19:16", "content": "I like your spirit!
", "rawMarkdown": "I like your spirit!", "votes": null }, { "id": "315336", "postDate": "04/17/2018 01:41:49", "content": "Hi,Oscar. I am using mask rcnn now. It got LB 0.590. It is hard to tell which is good. I would be very glad to see some public unet codes.
", "rawMarkdown": "Hi,Oscar. I am using mask rcnn now. It got LB 0.590. It is hard to tell which is good. I would be very glad to see some public unet codes.", "votes": null }, { "id": "315338", "postDate": "04/17/2018 01:43:32", "content": "Hi Heng. Thank you very much.I am using mask rcnn. I got 0.590.
", "rawMarkdown": "Hi Heng. Thank you very much.I am using mask rcnn. I got 0.590.", "votes": null }, { "id": "315342", "postDate": "04/17/2018 01:53:23", "content": "Thank you very much @Heng! We two fresh guys can got 0.55 finally because of your super code and idea! You are the best player in our mind!
", "rawMarkdown": "Thank you very much @Heng! We two fresh guys can got 0.55 finally because of your super code and idea! You are the best player in our mind!", "votes": null }, { "id": "315347", "postDate": "04/17/2018 02:03:18", "content": "That is pretty good.I use Masck-rcnn with resnext50 ,but only got 0.469.Could you show some detail about your model?
", "rawMarkdown": "That is pretty good.I use Masck-rcnn with resnext50 ,but only got 0.469.Could you show some detail about your model?", "votes": null }, { "id": "315358", "postDate": "04/17/2018 02:22:17", "content": "@Heng great inspiration ... learned a lot from you.
", "rawMarkdown": "Heng great inspiration ... learned a lot from you.", "votes": null }, { "id": "315362", "postDate": "04/17/2018 02:25:02", "content": "Thanks @Heng CherKeng! We learned a lot from your post!
", "rawMarkdown": "Thanks @Heng CherKeng! We learned a lot from your post!", "votes": null }, { "id": "315363", "postDate": "04/17/2018 02:25:34", "content": "Hi Cheng, thank you very much. I really learned a lot from this competition from you. Besides your share, I also appreciate your spirit!!
", "rawMarkdown": "Hi Cheng, thank you very much. I really learned a lot from this competition from you. Besides your share, I also appreciate your spirit!!", "votes": null }, { "id": "315364", "postDate": "04/17/2018 02:27:30", "content": "@Heng,Thanks for your sharing in this competition!
", "rawMarkdown": "Heng,Thanks for your sharing in this competition!", "votes": null }, { "id": "315369", "postDate": "04/17/2018 02:46:47", "content": "Congrats to all the winners!!!
", "rawMarkdown": "Congrats to all the winners!!!", "votes": null }, { "id": "315370", "postDate": "04/17/2018 02:48:31", "content": "@Heng, thank you for your posts, we learned a lot!
", "rawMarkdown": "Heng, thank you for your posts, we learned a lot!", "votes": null }, { "id": "315372", "postDate": "04/17/2018 02:49:02", "content": "Thank you for sharing your work @ Heng CherKeng. I learned a lot in this competition. This is my first time participating a kaggle competition, building and training a neural network, learning how to do image segmentation....... What I have learned from you and this community is 100x more than school classes and online courses.
", "rawMarkdown": "Thank you for sharing your work @ Heng CherKeng. I learned a lot in this competition. This is my first time participating a kaggle competition, building and training a neural network, learning how to do image segmentation....... What I have learned from you and this community is 100x more than school classes and online courses.", "votes": null }, { "id": "315375", "postDate": "04/17/2018 03:06:17", "content": "Thank you for sharing your work Heng. I hope your will make your mask rcnn implementation a proper library in the future. Also I should thank Waleed for his mask rcnn code which I used in the competition.
", "rawMarkdown": "Thank you for sharing your work Heng. I hope your will make your mask rcnn implementation a proper library in the future. Also I should thank Waleed for his mask rcnn code which I used in the competition.", "votes": null }, { "id": "315377", "postDate": "04/17/2018 03:06:50", "content": "Really amazing score for a single Mask-RCNN. Will you share your code later?
\n\n\n", "rawMarkdown": "Really amazing score for a single Mask-RCNN. Will you share your code later?\n\n> **Igor Krashenyi wrote**\n> \n> > we've managed to get 0.585 with a single Mask-RCNN, but with late submission :( \n> \n> http://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png", "votes": null }, { "id": "315381", "postDate": "04/17/2018 03:11:33", "content": "Igor Krashenyi wrote
\n \n\n\n \n \nwe've managed to get 0.585 with a single Mask-RCNN, but with late submission :(
\n
Thanks for lots of really inspiring posts!
", "rawMarkdown": "Thanks for lots of really inspiring posts!", "votes": null }, { "id": "315385", "postDate": "04/17/2018 03:28:18", "content": "Thanks for all contribution you have done to this competition, like @Russ W said, you embody the truth spirit of Kaggle. I can't wait to follow you in the next competition : ）. Thansk again!!
", "rawMarkdown": "Thanks for all contribution you have done to this competition, like @Russ W said, you embody the truth spirit of Kaggle. I can't wait to follow you in the next competition : ）. Thansk again!!", "votes": null }, { "id": "315391", "postDate": "04/17/2018 03:46:47", "content": "Mine is a single Mask-RCNN 0.591, I had another submission which I did not select have higher score but that does not count, but shows the potential.
", "rawMarkdown": "Mine is a single Mask-RCNN 0.591, I had another submission which I did not select have higher score but that does not count, but shows the potential.", "votes": null }, { "id": "315398", "postDate": "04/17/2018 04:06:15", "content": "@Heng Thank you so much for your contributions to this competition. You provide really good instructions and inspirations for most of us to participate in this contest. I have learnt a lot from your code and discussions. I think @Heng's contribution at least deserves a gold medal.
\n\nSo I also feel curious about the medal problem. As more than 3600+ teams participate in stage 1 but only 800+ teams submit results in stage 2. It seems that private LB is calculated based on stage 2 number. But will the final percentage be calculated based on the total number(3600+) or number in stage 2(800+)?
", "rawMarkdown": "Heng Thank you so much for your contributions to this competition. You provide really good instructions and inspirations for most of us to participate in this contest. I have learnt a lot from your code and discussions. I think @Heng's contribution at least deserves a gold medal. \n\nSo I also feel curious about the medal problem. As more than 3600+ teams participate in stage 1 but only 800+ teams submit results in stage 2. It seems that private LB is calculated based on stage 2 number. But will the final percentage be calculated based on the total number(3600+) or number in stage 2(800+)?", "votes": null }, { "id": "315401", "postDate": "04/17/2018 04:11:19", "content": "@Gangadhar, that is awesome! I also used Mask-RCNN with 22 epochs and got 0.489.\ncould you share your code and learning schedule on github ? \nMine is here but it is not fully polished yet. https://github.com/mpsampat/dsb_2018
", "rawMarkdown": "Gangadhar, that is awesome! I also used Mask-RCNN with 22 epochs and got 0.489.\ncould you share your code and learning schedule on github ? \nMine is here but it is not fully polished yet. https://github.com/mpsampat/dsb_2018", "votes": null }, { "id": "315413", "postDate": "04/17/2018 04:31:09", "content": "Hi Mehul. 0.489 is so good. Can you give me some details about it?
", "rawMarkdown": "Hi Mehul. 0.489 is so good. Can you give me some details about it?", "votes": null }, { "id": "315414", "postDate": "04/17/2018 04:37:21", "content": "Hi Zhiran. I am very happy to share my solution. If you don't mind, this is my wechat account: My-dear-lacus, we can discuss via wechat.It is much more convenient to discuss here.
", "rawMarkdown": "Hi Zhiran. I am very happy to share my solution. If you don't mind, this is my wechat account: My-dear-lacus, we can discuss via wechat.It is much more convenient to discuss here.", "votes": null }, { "id": "315429", "postDate": "04/17/2018 05:09:07", "content": "Can we talk in wechat too? I also use mask rcnn. I'm curious about ensemble & post process for mask rcnn.
", "rawMarkdown": "Can we talk in wechat too? I also use mask rcnn. I'm curious about ensemble & post process for mask rcnn.", "votes": null }, { "id": "315430", "postDate": "04/17/2018 05:13:29", "content": "Congrats to the winners and perhaps we can get some insights on your strategies?\nThank you.
", "rawMarkdown": "Congrats to the winners and perhaps we can get some insights on your strategies?\nThank you.", "votes": null }, { "id": "315435", "postDate": "04/17/2018 05:30:32", "content": "@heng Thanks for your posts, I've definitely learnt alot from you!.
", "rawMarkdown": "heng Thanks for your posts, I've definitely learnt alot from you!.", "votes": null }, { "id": "315437", "postDate": "04/17/2018 05:32:28", "content": "@Heng, I also think you deserve a gold medal too. I am a newbie in image research. Your ideas inspired me a lot. Many thanks!
", "rawMarkdown": "Heng, I also think you deserve a gold medal too. I am a newbie in image research. Your ideas inspired me a lot. Many thanks!", "votes": null }, { "id": "315440", "postDate": "04/17/2018 05:44:42", "content": "Based on the number of participants at the first stage
", "rawMarkdown": "Based on the number of participants at the first stage", "votes": null }, { "id": "315447", "postDate": "04/17/2018 06:06:46", "content": "@Heng, you inspired us to try different ideas and we often discussed your ideas at our coffee breaks (within the same team of course ;) ) You convinced me to split our models to two, one for grey and the other for colour, because you said something like \"The common trait of colour images and grey images is only the boundary. If they're trained together, it'd be hard for the neural net to learn the features inside the nucleus.\" Besides, we added speckle noise to our data augmentation after having seen your idea in the final 3 days. That made a single U-net a lot more robust too. You also convinced us to try mask RCNN (including matterport's) though we didn't make it to work as well as a single U-net and discarded the idea. As @Russ said, you embodied the true spirit of Kaggle. Sharing your high-LB-code during a competition is altruistic and respectful. I look forward to seeing you in the same competition again.
", "rawMarkdown": "Heng, you inspired us to try different ideas and we often discussed your ideas at our coffee breaks (within the same team of course ;) ) You convinced me to split our models to two, one for grey and the other for colour, because you said something like \"The common trait of colour images and grey images is only the boundary. If they're trained together, it'd be hard for the neural net to learn the features inside the nucleus.\" Besides, we added speckle noise to our data augmentation after having seen your idea in the final 3 days. That made a single U-net a lot more robust too. You also convinced us to try mask RCNN (including matterport's) though we didn't make it to work as well as a single U-net and discarded the idea. As @Russ said, you embodied the true spirit of Kaggle. Sharing your high-LB-code during a competition is altruistic and respectful. I look forward to seeing you in the same competition again.", "votes": null }, { "id": "315451", "postDate": "04/17/2018 06:12:14", "content": "Congrats to everyone
", "rawMarkdown": "Congrats to everyone", "votes": null }, { "id": "315474", "postDate": "04/17/2018 06:39:08", "content": "Congratulations to all. @heng, Thank you for your sharing so many thoughts and excellent code.
", "rawMarkdown": "Congratulations to all. @heng, Thank you for your sharing so many thoughts and excellent code.", "votes": null }, { "id": "315486", "postDate": "04/17/2018 07:07:24", "content": "Thank you @Heng and @CPMP for all your shared ideas and contributions! I learned a lot through this competition. And I still hope for a small rounding error to give Heng his deserved gold medal ;)
", "rawMarkdown": "Thank you @Heng and @CPMP for all your shared ideas and contributions! I learned a lot through this competition. And I still hope for a small rounding error to give Heng his deserved gold medal ;)", "votes": null }, { "id": "315503", "postDate": "04/17/2018 07:32:14", "content": "Nice one @heng. You should start a fan club. I'd join it.
\n\nWhat are we going to do next? The Kaggle competition list is looking a bit thin!
", "rawMarkdown": "Nice one @heng. You should start a fan club. I'd join it.\n\nWhat are we going to do next? The Kaggle competition list is looking a bit thin!", "votes": null }, { "id": "315505", "postDate": "04/17/2018 07:35:21", "content": "Thanks Heng for all the inside you shared !
", "rawMarkdown": "Thanks Heng for all the inside you shared !", "votes": null }, { "id": "315582", "postDate": "04/17/2018 10:00:56", "content": "Congratulations to the winners!!! I look forward to reading about your solutions soon.
\n\n@Heng thanks for all your contributions.
\n\nThanks @Jakub and @Kamil from Neptune ML for your contributions.
\n\nThanks everyone else for making this a lively competition. See you in the next one.
\n\nHappy Kaggling!!!
", "rawMarkdown": "Congratulations to the winners!!! I look forward to reading about your solutions soon.\n\n@Heng thanks for all your contributions.\n\nThanks @Jakub and @Kamil from Neptune ML for your contributions.\n\nThanks everyone else for making this a lively competition. See you in the next one.\n\nHappy Kaggling!!!", "votes": null }, { "id": "315678", "postDate": "04/17/2018 13:14:59", "content": "great!!!
", "rawMarkdown": "great!!!", "votes": null }, { "id": "315689", "postDate": "04/17/2018 13:34:35", "content": "Thx ! Learnt a lot from this competition.
", "rawMarkdown": "Thx ! Learnt a lot from this competition.", "votes": null }, { "id": "315770", "postDate": "04/17/2018 15:17:40", "content": "Got 0.56 with Mask-RCNN
", "rawMarkdown": "Got 0.56 with Mask-RCNN", "votes": null }, { "id": "315983", "postDate": "04/17/2018 23:35:45", "content": "@JamesEverard
\n\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?
\n\nwill let you know when they are setup. should take a few week
", "rawMarkdown": "JamesEverard\n\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?\n\nwill let you know when they are setup. should take a few week", "votes": null }, { "id": "316000", "postDate": "04/18/2018 01:12:27", "content": "Discourse https://github.com/discourse/discourse
\n\n\n", "rawMarkdown": "Discourse https://github.com/discourse/discourse\n\n> **Heng CherKeng wrote**\n> \n> > @JamesEverard\n> \n> I am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?\n> \n> will let you know when they are setup. should take a few week", "votes": null }, { "id": "316035", "postDate": "04/18/2018 03:52:11", "content": "Heng CherKeng wrote
\n \n\n\n \n@JamesEverard
\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?
\n \nwill let you know when they are setup. should take a few week
\n
@heng thanks for all your contributions here and in other competitions, you are truly generous and embody a good spirit - always looking to new ideas and to share with others. If there were a Nobel Prize for Kaggle you should get it! Glad to hear about your upcoming blog. Was going to suggest in your next competition to have a pinned thread for Heng's content!
\nAll the best and congrats to the winners and contributors.
from there name I can here @Deep Retina used retina net maybe - I am interesting to here there solution
", "rawMarkdown": "from there name I can here @Deep Retina used retina net maybe - I am interesting to here there solution", "votes": null }, { "id": "316119", "postDate": "04/18/2018 08:09:29", "content": "Great idea,
\n\nI'm happy to write some data related kernels for you so you can focus on the cool stuff.
", "rawMarkdown": "Great idea,\n\nI'm happy to write some data related kernels for you so you can focus on the cool stuff.", "votes": null }, { "id": "316133", "postDate": "04/18/2018 08:32:18", "content": "Thank you
", "rawMarkdown": "Thank you", "votes": null }, { "id": "316135", "postDate": "04/18/2018 08:33:32", "content": "@Heng
\n\nYes, v good idea. It will be nice to have top solutions to past competitions. There will also be much more scope for discussion once we start to implement them.
", "rawMarkdown": "Heng\n\nYes, v good idea. It will be nice to have top solutions to past competitions. There will also be much more scope for discussion once we start to implement them.", "votes": null } ], "comments": [ { "id": 315285, "author_name": "watts2", "author_url": "", "post_date": "04/17/2018 00:44:42", "content": "Congratulations to all winners! Well done!
\n\nSpecial thank you to @ Heng CherKeng for the excellent code, discussions and generous contributions. The competition would not be so interesting if it was not for these contributions.
", "votes": null, "replies": [] }, { "id": 315286, "author_name": "cpmpml", "author_url": "", "post_date": "04/17/2018 00:47:00", "content": "\n\n\nSo UNet Vs Mask-rcnn, which is better?
\n
Maybe a combination of them?
\n\nFrom my computation based on stage1 participation, you're one short of gold medal, which is a pity given all you shared. I can't say I hope someone will be removed, but you really deserve a gold medal here.
", "votes": null, "replies": [ { "id": 315288, "author_name": "hengck23", "author_url": "", "post_date": "04/17/2018 00:48:35", "content": "@CPMP no worry. there is always the next kaggle challenge!
", "votes": null, "replies": [] }, { "id": 315326, "author_name": "outliar", "author_url": "", "post_date": "04/17/2018 01:19:16", "content": "I like your spirit!
", "votes": null, "replies": [] }, { "id": 315437, "author_name": "chiaosun", "author_url": "", "post_date": "04/17/2018 05:32:28", "content": "@Heng, I also think you deserve a gold medal too. I am a newbie in image research. Your ideas inspired me a lot. Many thanks!
", "votes": null, "replies": [] }, { "id": 315486, "author_name": "florianhuber", "author_url": "", "post_date": "04/17/2018 07:07:24", "content": "Thank you @Heng and @CPMP for all your shared ideas and contributions! I learned a lot through this competition. And I still hope for a small rounding error to give Heng his deserved gold medal ;)
", "votes": null, "replies": [] } ] }, { "id": 315292, "author_name": "pliptor", "author_url": "", "post_date": "04/17/2018 00:51:48", "content": "Has anyone tried something other than U-Net or Mask-RCNN and got a score above 0.500?
", "votes": null, "replies": [ { "id": 315336, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/17/2018 01:41:49", "content": "Hi,Oscar. I am using mask rcnn now. It got LB 0.590. It is hard to tell which is good. I would be very glad to see some public unet codes.
", "votes": null, "replies": [] } ] }, { "id": 315294, "author_name": "sasrdw", "author_url": "", "post_date": "04/17/2018 00:52:24", "content": "Thanks again Heng for really embodying the true spirit of Kaggle. My current impression is modifications of UNet are a little better for this problem--that's what we used. Allen just reported 0.582 private for his best Mask-RCNN model; anybody beat that? @CPMP I was dying to try and find a Mask-RCNN result to ensemble with our good UNet but never found one that passed comparable muster locally.
", "votes": null, "replies": [ { "id": 315308, "author_name": "igorkrashenyi", "author_url": "", "post_date": "04/17/2018 01:04:54", "content": "we've managed to get 0.585 with a single Mask-RCNN, but with late submission :(
\n\n", "votes": null, "replies": [] }, { "id": 315325, "author_name": "cpmpml", "author_url": "", "post_date": "04/17/2018 01:19:07", "content": "@Russ W, Thanks for sharing, can't wait for a more complete write up form your team!
", "votes": null, "replies": [] }, { "id": 315377, "author_name": "brianlzm", "author_url": "", "post_date": "04/17/2018 03:06:50", "content": "Really amazing score for a single Mask-RCNN. Will you share your code later?
\n\n\n", "votes": null, "replies": [] }, { "id": 315391, "author_name": "gangadhar", "author_url": "", "post_date": "04/17/2018 03:46:47", "content": "Igor Krashenyi wrote
\n \n\n\n \n \nwe've managed to get 0.585 with a single Mask-RCNN, but with late submission :(
\n
Mine is a single Mask-RCNN 0.591, I had another submission which I did not select have higher score but that does not count, but shows the potential.
", "votes": null, "replies": [] }, { "id": 315401, "author_name": "mpsampat", "author_url": "", "post_date": "04/17/2018 04:11:19", "content": "@Gangadhar, that is awesome! I also used Mask-RCNN with 22 epochs and got 0.489.\ncould you share your code and learning schedule on github ? \nMine is here but it is not fully polished yet. https://github.com/mpsampat/dsb_2018
", "votes": null, "replies": [] }, { "id": 315413, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/17/2018 04:31:09", "content": "Hi Mehul. 0.489 is so good. Can you give me some details about it?
", "votes": null, "replies": [] }, { "id": 315770, "author_name": "princerk", "author_url": "", "post_date": "04/17/2018 15:17:40", "content": "Got 0.56 with Mask-RCNN
", "votes": null, "replies": [] } ] }, { "id": 315338, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/17/2018 01:43:32", "content": "Hi Heng. Thank you very much.I am using mask rcnn. I got 0.590.
", "votes": null, "replies": [ { "id": 315347, "author_name": "hellopipu333", "author_url": "", "post_date": "04/17/2018 02:03:18", "content": "That is pretty good.I use Masck-rcnn with resnext50 ,but only got 0.469.Could you show some detail about your model?
", "votes": null, "replies": [] }, { "id": 315414, "author_name": "mdlszhengli", "author_url": "", "post_date": "04/17/2018 04:37:21", "content": "Hi Zhiran. I am very happy to share my solution. If you don't mind, this is my wechat account: My-dear-lacus, we can discuss via wechat.It is much more convenient to discuss here.
", "votes": null, "replies": [] }, { "id": 315429, "author_name": "brianlzm", "author_url": "", "post_date": "04/17/2018 05:09:07", "content": "Can we talk in wechat too? I also use mask rcnn. I'm curious about ensemble & post process for mask rcnn.
", "votes": null, "replies": [] } ] }, { "id": 315342, "author_name": "nishinihaha", "author_url": "", "post_date": "04/17/2018 01:53:23", "content": "Thank you very much @Heng! We two fresh guys can got 0.55 finally because of your super code and idea! You are the best player in our mind!
", "votes": null, "replies": [] }, { "id": 315358, "author_name": "aftabk", "author_url": "", "post_date": "04/17/2018 02:22:17", "content": "@Heng great inspiration ... learned a lot from you.
", "votes": null, "replies": [] }, { "id": 315362, "author_name": "panpan7967", "author_url": "", "post_date": "04/17/2018 02:25:02", "content": "Thanks @Heng CherKeng! We learned a lot from your post!
", "votes": null, "replies": [] }, { "id": 315363, "author_name": "zhen27", "author_url": "", "post_date": "04/17/2018 02:25:34", "content": "Hi Cheng, thank you very much. I really learned a lot from this competition from you. Besides your share, I also appreciate your spirit!!
", "votes": null, "replies": [] }, { "id": 315364, "author_name": "hellopipu333", "author_url": "", "post_date": "04/17/2018 02:27:30", "content": "@Heng,Thanks for your sharing in this competition!
", "votes": null, "replies": [] }, { "id": 315369, "author_name": "samratp", "author_url": "", "post_date": "04/17/2018 02:46:47", "content": "Congrats to all the winners!!!
", "votes": null, "replies": [] }, { "id": 315370, "author_name": "yanjunyang", "author_url": "", "post_date": "04/17/2018 02:48:31", "content": "@Heng, thank you for your posts, we learned a lot!
", "votes": null, "replies": [] }, { "id": 315372, "author_name": "chouchris", "author_url": "", "post_date": "04/17/2018 02:49:02", "content": "Thank you for sharing your work @ Heng CherKeng. I learned a lot in this competition. This is my first time participating a kaggle competition, building and training a neural network, learning how to do image segmentation....... What I have learned from you and this community is 100x more than school classes and online courses.
", "votes": null, "replies": [] }, { "id": 315375, "author_name": "akilaw", "author_url": "", "post_date": "04/17/2018 03:06:17", "content": "Thank you for sharing your work Heng. I hope your will make your mask rcnn implementation a proper library in the future. Also I should thank Waleed for his mask rcnn code which I used in the competition.
", "votes": null, "replies": [] }, { "id": 315381, "author_name": "mengtianjian", "author_url": "", "post_date": "04/17/2018 03:11:33", "content": "Thanks for lots of really inspiring posts!
", "votes": null, "replies": [] }, { "id": 315385, "author_name": "lincurry", "author_url": "", "post_date": "04/17/2018 03:28:18", "content": "Thanks for all contribution you have done to this competition, like @Russ W said, you embody the truth spirit of Kaggle. I can't wait to follow you in the next competition : ）. Thansk again!!
", "votes": null, "replies": [] }, { "id": 315398, "author_name": "haroldlee", "author_url": "", "post_date": "04/17/2018 04:06:15", "content": "@Heng Thank you so much for your contributions to this competition. You provide really good instructions and inspirations for most of us to participate in this contest. I have learnt a lot from your code and discussions. I think @Heng's contribution at least deserves a gold medal.
\n\nSo I also feel curious about the medal problem. As more than 3600+ teams participate in stage 1 but only 800+ teams submit results in stage 2. It seems that private LB is calculated based on stage 2 number. But will the final percentage be calculated based on the total number(3600+) or number in stage 2(800+)?
", "votes": null, "replies": [ { "id": 315440, "author_name": "nicolefinnie", "author_url": "", "post_date": "04/17/2018 05:44:42", "content": "Based on the number of participants at the first stage
", "votes": null, "replies": [] }, { "id": 315689, "author_name": "haroldlee", "author_url": "", "post_date": "04/17/2018 13:34:35", "content": "Thx ! Learnt a lot from this competition.
", "votes": null, "replies": [] } ] }, { "id": 315430, "author_name": "ishootlaser", "author_url": "", "post_date": "04/17/2018 05:13:29", "content": "Congrats to the winners and perhaps we can get some insights on your strategies?\nThank you.
", "votes": null, "replies": [] }, { "id": 315435, "author_name": "worulz", "author_url": "", "post_date": "04/17/2018 05:30:32", "content": "@heng Thanks for your posts, I've definitely learnt alot from you!.
", "votes": null, "replies": [] }, { "id": 315447, "author_name": "nicolefinnie", "author_url": "", "post_date": "04/17/2018 06:06:46", "content": "@Heng, you inspired us to try different ideas and we often discussed your ideas at our coffee breaks (within the same team of course ;) ) You convinced me to split our models to two, one for grey and the other for colour, because you said something like \"The common trait of colour images and grey images is only the boundary. If they're trained together, it'd be hard for the neural net to learn the features inside the nucleus.\" Besides, we added speckle noise to our data augmentation after having seen your idea in the final 3 days. That made a single U-net a lot more robust too. You also convinced us to try mask RCNN (including matterport's) though we didn't make it to work as well as a single U-net and discarded the idea. As @Russ said, you embodied the true spirit of Kaggle. Sharing your high-LB-code during a competition is altruistic and respectful. I look forward to seeing you in the same competition again.
", "votes": null, "replies": [] }, { "id": 315451, "author_name": "steelrose", "author_url": "", "post_date": "04/17/2018 06:12:14", "content": "Congrats to everyone
", "votes": null, "replies": [] }, { "id": 315474, "author_name": "zmingen", "author_url": "", "post_date": "04/17/2018 06:39:08", "content": "Congratulations to all. @heng, Thank you for your sharing so many thoughts and excellent code.
", "votes": null, "replies": [] }, { "id": 315503, "author_name": "jinkos", "author_url": "", "post_date": "04/17/2018 07:32:14", "content": "Nice one @heng. You should start a fan club. I'd join it.
\n\nWhat are we going to do next? The Kaggle competition list is looking a bit thin!
", "votes": null, "replies": [ { "id": 315983, "author_name": "hengck23", "author_url": "", "post_date": "04/17/2018 23:35:45", "content": "@JamesEverard
\n\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?
\n\nwill let you know when they are setup. should take a few week
", "votes": null, "replies": [] }, { "id": 316000, "author_name": "jamesrequa", "author_url": "", "post_date": "04/18/2018 01:12:27", "content": "Discourse https://github.com/discourse/discourse
\n\n\n", "votes": null, "replies": [] }, { "id": 316035, "author_name": "something4kag", "author_url": "", "post_date": "04/18/2018 03:52:11", "content": "Heng CherKeng wrote
\n \n\n\n \n@JamesEverard
\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?
\n \nwill let you know when they are setup. should take a few week
\n
@heng thanks for all your contributions here and in other competitions, you are truly generous and embody a good spirit - always looking to new ideas and to share with others. If there were a Nobel Prize for Kaggle you should get it! Glad to hear about your upcoming blog. Was going to suggest in your next competition to have a pinned thread for Heng's content!
\nAll the best and congrats to the winners and contributors.
Great idea,
\n\nI'm happy to write some data related kernels for you so you can focus on the cool stuff.
", "votes": null, "replies": [] }, { "id": 316135, "author_name": "watts2", "author_url": "", "post_date": "04/18/2018 08:33:32", "content": "@Heng
\n\nYes, v good idea. It will be nice to have top solutions to past competitions. There will also be much more scope for discussion once we start to implement them.
", "votes": null, "replies": [] } ] }, { "id": 315505, "author_name": "tuanphuc", "author_url": "", "post_date": "04/17/2018 07:35:21", "content": "Thanks Heng for all the inside you shared !
", "votes": null, "replies": [] }, { "id": 315582, "author_name": "sheriytm", "author_url": "", "post_date": "04/17/2018 10:00:56", "content": "Congratulations to the winners!!! I look forward to reading about your solutions soon.
\n\n@Heng thanks for all your contributions.
\n\nThanks @Jakub and @Kamil from Neptune ML for your contributions.
\n\nThanks everyone else for making this a lively competition. See you in the next one.
\n\nHappy Kaggling!!!
", "votes": null, "replies": [] }, { "id": 315678, "author_name": "anishjosh", "author_url": "", "post_date": "04/17/2018 13:14:59", "content": "great!!!
", "votes": null, "replies": [] }, { "id": 316065, "author_name": "aviwolfson", "author_url": "", "post_date": "04/18/2018 05:39:51", "content": "from there name I can here @Deep Retina used retina net maybe - I am interesting to here there solution
", "votes": null, "replies": [] }, { "id": 316133, "author_name": "andre1233", "author_url": "", "post_date": "04/18/2018 08:32:18", "content": "Thank you
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "315281": "Congrats to @[ods.ai] topcoders, @jacobkie, @Deep Retina, @Nuclear Vision, @Inom Mirzaev for being the top 5 and all kagglers who had completed the challenge.\n\nIt has been a long and tough challenge. All has done well! I think the scores of the top 5 are amazing and their models did generalise very well. Can't wait to hear more about their solutions!\n\nSo UNet Vs Mask-rcnn, which is better?", "315285": "Congratulations to all winners! Well done!\n\nSpecial thank you to @ Heng CherKeng for the excellent code, discussions and generous contributions. The competition would not be so interesting if it was not for these contributions.", "315286": "> So UNet Vs Mask-rcnn, which is better?\n\nMaybe a combination of them?\n\nFrom my computation based on stage1 participation, you're one short of gold medal, which is a pity given all you shared. I can't say I hope someone will be removed, but you really deserve a gold medal here.", "315288": "CPMP no worry. there is always the next kaggle challenge!", "315292": "Has anyone tried something other than U-Net or Mask-RCNN and got a score above 0.500?", "315294": "Thanks again Heng for really embodying the true spirit of Kaggle. My current impression is modifications of UNet are a little better for this problem--that's what we used. Allen just reported 0.582 private for his best Mask-RCNN model; anybody beat that? @CPMP I was *dying* to try and find a Mask-RCNN result to ensemble with our good UNet but never found one that passed comparable muster locally.", "315308": "we've managed to get 0.585 with a single Mask-RCNN, but with late submission :( \n\nhttp://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png", "315325": "Russ W, Thanks for sharing, can't wait for a more complete write up form your team!", "315326": "I like your spirit!", "315336": "Hi,Oscar. I am using mask rcnn now. It got LB 0.590. It is hard to tell which is good. I would be very glad to see some public unet codes.", "315338": "Hi Heng. Thank you very much.I am using mask rcnn. I got 0.590.", "315342": "Thank you very much @Heng! We two fresh guys can got 0.55 finally because of your super code and idea! You are the best player in our mind!", "315347": "That is pretty good.I use Masck-rcnn with resnext50 ,but only got 0.469.Could you show some detail about your model?", "315358": "Heng great inspiration ... learned a lot from you.", "315362": "Thanks @Heng CherKeng! We learned a lot from your post!", "315363": "Hi Cheng, thank you very much. I really learned a lot from this competition from you. Besides your share, I also appreciate your spirit!!", "315364": "Heng,Thanks for your sharing in this competition!", "315369": "Congrats to all the winners!!!", "315370": "Heng, thank you for your posts, we learned a lot!", "315372": "Thank you for sharing your work @ Heng CherKeng. I learned a lot in this competition. This is my first time participating a kaggle competition, building and training a neural network, learning how to do image segmentation....... What I have learned from you and this community is 100x more than school classes and online courses.", "315375": "Thank you for sharing your work Heng. I hope your will make your mask rcnn implementation a proper library in the future. Also I should thank Waleed for his mask rcnn code which I used in the competition.", "315377": "Really amazing score for a single Mask-RCNN. Will you share your code later?\n\n> **Igor Krashenyi wrote**\n> \n> > we've managed to get 0.585 with a single Mask-RCNN, but with late submission :( \n> \n> http://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png", "315381": "Thanks for lots of really inspiring posts!", "315385": "Thanks for all contribution you have done to this competition, like @Russ W said, you embody the truth spirit of Kaggle. I can't wait to follow you in the next competition : ）. Thansk again!!", "315391": "Mine is a single Mask-RCNN 0.591, I had another submission which I did not select have higher score but that does not count, but shows the potential.", "315398": "Heng Thank you so much for your contributions to this competition. You provide really good instructions and inspirations for most of us to participate in this contest. I have learnt a lot from your code and discussions. I think @Heng's contribution at least deserves a gold medal. \n\nSo I also feel curious about the medal problem. As more than 3600+ teams participate in stage 1 but only 800+ teams submit results in stage 2. It seems that private LB is calculated based on stage 2 number. But will the final percentage be calculated based on the total number(3600+) or number in stage 2(800+)?", "315401": "Gangadhar, that is awesome! I also used Mask-RCNN with 22 epochs and got 0.489.\ncould you share your code and learning schedule on github ? \nMine is here but it is not fully polished yet. https://github.com/mpsampat/dsb_2018", "315413": "Hi Mehul. 0.489 is so good. Can you give me some details about it?", "315414": "Hi Zhiran. I am very happy to share my solution. If you don't mind, this is my wechat account: My-dear-lacus, we can discuss via wechat.It is much more convenient to discuss here.", "315429": "Can we talk in wechat too? I also use mask rcnn. I'm curious about ensemble & post process for mask rcnn.", "315430": "Congrats to the winners and perhaps we can get some insights on your strategies?\nThank you.", "315435": "heng Thanks for your posts, I've definitely learnt alot from you!.", "315437": "Heng, I also think you deserve a gold medal too. I am a newbie in image research. Your ideas inspired me a lot. Many thanks!", "315440": "Based on the number of participants at the first stage", "315447": "Heng, you inspired us to try different ideas and we often discussed your ideas at our coffee breaks (within the same team of course ;) ) You convinced me to split our models to two, one for grey and the other for colour, because you said something like \"The common trait of colour images and grey images is only the boundary. If they're trained together, it'd be hard for the neural net to learn the features inside the nucleus.\" Besides, we added speckle noise to our data augmentation after having seen your idea in the final 3 days. That made a single U-net a lot more robust too. You also convinced us to try mask RCNN (including matterport's) though we didn't make it to work as well as a single U-net and discarded the idea. As @Russ said, you embodied the true spirit of Kaggle. Sharing your high-LB-code during a competition is altruistic and respectful. I look forward to seeing you in the same competition again.", "315451": "Congrats to everyone", "315474": "Congratulations to all. @heng, Thank you for your sharing so many thoughts and excellent code.", "315486": "Thank you @Heng and @CPMP for all your shared ideas and contributions! I learned a lot through this competition. And I still hope for a small rounding error to give Heng his deserved gold medal ;)", "315503": "Nice one @heng. You should start a fan club. I'd join it.\n\nWhat are we going to do next? The Kaggle competition list is looking a bit thin!", "315505": "Thanks Heng for all the inside you shared !", "315582": "Congratulations to the winners!!! I look forward to reading about your solutions soon.\n\n@Heng thanks for all your contributions.\n\nThanks @Jakub and @Kamil from Neptune ML for your contributions.\n\nThanks everyone else for making this a lively competition. See you in the next one.\n\nHappy Kaggling!!!", "315678": "great!!!", "315689": "Thx ! Learnt a lot from this competition.", "315770": "Got 0.56 with Mask-RCNN", "315983": "JamesEverard\n\nI am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?\n\nwill let you know when they are setup. should take a few week", "316000": "Discourse https://github.com/discourse/discourse\n\n> **Heng CherKeng wrote**\n> \n> > @JamesEverard\n> \n> I am setting up a blog and discussion to repeat pass kaggle top solution from scratch. I will be reimplementing topcoders solution for DSB 2018. I will be using medium for blogging and is looking for good for a good forum platform. Does anyone know what is the forum platform that powers fast.ai or discuss.pytorch.com?\n> \n> will let you know when they are setup. should take a few week", "316035": "heng thanks for all your contributions here and in other competitions, you are truly generous and embody a good spirit - always looking to new ideas and to share with others. If there were a Nobel Prize for Kaggle you should get it! Glad to hear about your upcoming blog. Was going to suggest in your next competition to have a pinned thread for Heng's content! \nAll the best and congrats to the winners and contributors.", "316065": "from there name I can here @Deep Retina used retina net maybe - I am interesting to here there solution", "316119": "Great idea,\n\nI'm happy to write some data related kernels for you so you can focus on the cool stuff.", "316133": "Thank you", "316135": "Heng\n\nYes, v good idea. It will be nice to have top solutions to past competitions. There will also be much more scope for discussion once we start to implement them." }, "source": "meta" }![http://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png](\"http://i.piccy.info/i9/925921ed2beb0474f50f8c7f9b168be1/1523926942/33763/1237800/Screen_Shot_2018_04_17_at_03_58_19.png\"){kind=link}