\n
![\"rank\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2081456%2F59d81975ca1ac9373e390e43185c6824%2F-6448bed0f1d976b1.jpg?generation=1574214746197536&alt=media\")
\nSong is a friend of us actully.I mean she is a fabulous guy.She got a great insight upon this competition. It's my very first CV competition in Kaggle,APTOS blindess is so hard so I gave it up very early,but I 've been doing my best in this segmentation comp.
\n\n#\n\nPity? yes,I could have a chance to get a silver in the end,there was five kernels running KFold last day(UTC+8) which I think will eventually give a boost to my result,but unfortunately,three of them were dead due to the time limitation ,so I couldn't make full use of all my resources,but I got to say this is part of this competition,like Yakovlev\n saied here ,now that you choose to be a part of this comp,you should accept this unfairness,BTW,I'm a big fan of him and Chris,their insights of these competitions are so great.I learned a lot from them personally.
\n\n#\n\nWhat I learned from this comp is that you need to do a handful of job to get this done,I tried a lot of combinations of different encoders and decoders,and it turned out that efficientnet has a better performance, and due to the GPU time limitation .my teammate could not add attention method to the decoder and for sure this can give a boost.And it's drawing to the deadline when I tried Kfold.The reason y I can't try it earlier is the same thing due to the time limitation,so I have to use 5 kernels to save 5 models, and do ensemble later.Still got a lot of work to do in the end ,haven't found a good combination of different classifier that can boost my score. But for sure ,this can give a boost
\n\n#\n\nFinally,glad to win an 'expert' title in this wonderful community,you guys are just so Amazingggg!!!!
", "messages": [ { "id": "676146", "postDate": "11/19/2019 01:47:52", "content": "It's my very first CV competition in Kaggle,APTOS blindess is so hard so I gave it up very early,but I 've been doing my best in this segmentation comp.
\n\n#\n\nPity? yes,I could have a chance to get a silver in the end,there was five kernels running KFold last day(UTC+8) which I think will eventually give a boost to my result,but unfortunately,three of them were dead due to the time limitation ,so I couldn't make full use of all my resources,but I got to say this is part of this competition,like Yakovlev\n saied here ,now that you choose to be a part of this comp,you should accept this unfairness,BTW,I'm a big fan of him and Chris,their insights of these competitions are so great.I learned a lot from them personally.
\n\n#\n\nWhat I learned from this comp is that you need to do a handful of job to get this done,I tried a lot of combinations of different encoders and decoders,and it turned out that efficientnet has a better performance, and due to the GPU time limitation .my teammate could not add attention method to the decoder and for sure this can give a boost.And it's drawing to the deadline when I tried Kfold.The reason y I can't try it earlier is the same thing due to the time limitation,so I have to use 5 kernels to save 5 models, and do ensemble later.Still got a lot of work to do in the end ,haven't found a good combination of different classifier that can boost my score. But for sure ,this can give a boost
\n\n#\n\nFinally,glad to win an 'expert' title in this wonderful community,you guys are just so Amazingggg!!!!
", "rawMarkdown": "It's my very first CV competition in Kaggle,APTOS blindess is so hard so I gave it up very early,but I 've been doing my best in this segmentation comp. \n##########################################\nPity? yes,I could have a chance to get a silver in the end,there was five kernels running KFold last day(UTC+8) which I think will eventually give a boost to my result,but unfortunately,three of them were dead due to the time limitation ,so I couldn't make full use of all my resources,but I got to say this is part of this competition,like Yakovlev\n saied [here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/117016#latest-674872) ,now that you choose to be a part of this comp,you should accept this unfairness,BTW,I'm a big fan of him and Chris,their insights of these competitions are so great.I learned a lot from them personally.\n##########################################\nWhat I learned from this comp is that you need to do a handful of job to get this done,I tried a lot of combinations of different encoders and decoders,and it turned out that efficientnet has a better performance, and due to the GPU time limitation .my teammate could not add attention method to the decoder and for sure this can give a boost.And it's drawing to the deadline when I tried Kfold.The reason y I can't try it earlier is the same thing due to the time limitation,so I have to use 5 kernels to save 5 models, and do ensemble later.Still got a lot of work to do in the end ,haven't found a good combination of different classifier that can boost my score. But for sure ,this can give a boost\n##########################################\nFinally,glad to win an 'expert' title in this wonderful community,you guys are just so Amazingggg!!!!", "votes": null }, { "id": "676156", "postDate": "11/19/2019 01:58:48", "content": "ouch ! you just missed one silver .I know that feeling . But for what its worth , i think you are not really a \"Overfit Queen \" .. you jumped up 31 places :)
", "rawMarkdown": "ouch ! you just missed one silver .I know that feeling . But for what its worth , i think you are not really a \"Overfit Queen \" .. you jumped up 31 places :)", "votes": null }, { "id": "676159", "postDate": "11/19/2019 02:03:11", "content": "Yeah,I named myself this one cuz I got an over one thousand drop in the Lanl earthquake——my first round in this community,yet so hard to deal with time series data
", "rawMarkdown": "Yeah,I named myself this one cuz I got an over one thousand drop in the Lanl earthquake——my first round in this community,yet so hard to deal with time series data", "votes": null }, { "id": "676815", "postDate": "11/19/2019 14:23:46", "content": "You were just one step ahead of me when competition ended, why your team name removed from leaderboard?\nMy Suggestion to train a model if you don't know whether it will be completed in 9 hours.\n1. Save the states of model, optimizer and scheduler after each epoch.\n2. Save the states of model, optimizer and scheduler of best model till that epoch.\n3. write your log in a plain text file in unbuffered mode, since kernel logs won't be accessible if your kernel goes dead.
\n\nif your kernel becomes dead due to time expiration you will still have above 3 files and based on logs you can decides whether you can further trained it or not. if you wants to train it for few more epoch just load the weights of last epoch of your model and states of optimizer and scheduler, and start to train it further.
", "rawMarkdown": "You were just one step ahead of me when competition ended, why your team name removed from leaderboard?\nMy Suggestion to train a model if you don't know whether it will be completed in 9 hours.\n1. Save the states of model, optimizer and scheduler after each epoch.\n2. Save the states of model, optimizer and scheduler of best model till that epoch.\n3. write your log in a plain text file in unbuffered mode, since kernel logs won't be accessible if your kernel goes dead.\n\nif your kernel becomes dead due to time expiration you will still have above 3 files and based on logs you can decides whether you can further trained it or not. if you wants to train it for few more epoch just load the weights of last epoch of your model and states of optimizer and scheduler, and start to train it further.", "votes": null }, { "id": "677290", "postDate": "11/20/2019 02:11:46", "content": "Firstly congrats on the silver. Also we are so appreciated for your advice,will give it a try in our next competition.
\n\nWe are not able to survive the LB cleaning,yet we do not deseve it because we shared are results with our friends in the other team who should be merged to us before the merging deadline and broke the rules I was not familiar with before.Yet we don't have a chance to do this again.
\n\nHere is our ranking before the LB cleaning
\n\nSong is a friend of us actully.I mean she is a fabulous guy.She got a great insight upon this competition.
Here are parts of our proofs for this competition.I mean we stand as a team to shed sweat,paying as much as effort as you guys,working hard trying to find the solution.
\n![\"effort\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2081456%2Fb9214ec1a3f8849cf01da33cde16da31%2Feffort.jpg?generation=1574215256179112&alt=media\")
\n![\"effort\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2081456%2Fa410a880090cb45c9df2217639ed5619%2Fproject.jpg?generation=1574215320518936&alt=media\")
\n![\"signal\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2081456%2Fc0a19522b4191a1ff62543ab840a6cae%2Fsignal.jpg?generation=1574215576907071&alt=media\")
Basically, our method was based on @gogo827jz kernel And thx him for sharing his wonderful ideas
\n\nTo make it clear,we are not that kind of cheating guys,just made a mistake which should not be made in this community.
\n\nI do not feel any pity now cuz I've been there.And glad to learn and share informations with all you guys,and I will continue being humble,using knowledge to arm myself,and sharing everything I've got with you cuz I've learned so much from this community.
", "rawMarkdown": "Firstly congrats on the silver. Also we are so appreciated for your advice,will give it a try in our next competition. \n\nWe are not able to survive the LB cleaning,yet we do not deseve it because we shared are results with our friends in the other team who should be merged to us before the merging deadline and broke the rules I was not familiar with before.Yet we don't have a chance to do this again. \n\nHere is our ranking before the LB cleaning \n\nSong is a friend of us actully.I mean she is a fabulous guy.She got a great insight upon this competition. \n\nHere are parts of our proofs for this competition.I mean we stand as a team to shed sweat,paying as much as effort as you guys,working hard trying to find the solution. \n \n \n\n\nBasically, our method was based on @gogo827jz [kernel](https://www.kaggle.com/gogo827jz/resunet-keras-with-some-new-ideas) And thx him for sharing his wonderful ideas\n\nTo make it clear,we are not that kind of cheating guys,just made a mistake which should not be made in this community. \n\nI do not feel any pity now cuz I've been there.And glad to learn and share informations with all you guys,and I will continue being humble,using knowledge to arm myself,and sharing everything I've got with you cuz I've learned so much from this community.", "votes": null }, { "id": "677335", "postDate": "11/20/2019 03:55:12", "content": ">now that you choose to be a part of this comp,you should accept this unfairness
\n\nI don't get it, where is the unfairness in this competition?
", "rawMarkdown": ">now that you choose to be a part of this comp,you should accept this unfairness\n\nI don't get it, where is the unfairness in this competition?", "votes": null }, { "id": "677346", "postDate": "11/20/2019 04:03:35", "content": "Its really sad to hear this!!! Best of luck for next competitions.\nI have one question, if you can answer it will be a valuable knowledge for me.\nwhat is attention method to the decoder?
", "rawMarkdown": "Its really sad to hear this!!! Best of luck for next competitions.\nI have one question, if you can answer it will be a valuable knowledge for me.\nwhat is attention method to the decoder?", "votes": null }, { "id": "677355", "postDate": "11/20/2019 04:16:54", "content": "I use google colab since I don't have my own GPU. Since colab can goes off any time and therefore to prevent any loss of computation, I use follwong code to save my models, hope it will be help full to you.
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1490082%2F52fa57627f61515ad2629b33733546b3%2F1.png?generation=1574223247549954&alt=media\")
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1490082%2F548e9bff5802c8199a02ef9075f20e10%2F2.png?generation=1574223406819095&alt=media\")
Wow,that's so nice of you.That code is so wonderful,and I'm gonna write it down to my notes which can check the best model to save according to the cv score with pytorch.That's what I need for now,right in time.BTW，here is my answer for you question above,attention method is explained here and the code corresponding to this is \n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2081456%2F1abd4f6c505565cb6ebad074a42cdc55%2Fattention.jpg?generation=1574231679547509&alt=media\")
I found this in a blog written in mainland China if you are interested here
", "rawMarkdown": "Wow,that's so nice of you.That code is so wonderful,and I'm gonna write it down to my notes which can check the best model to save according to the cv score with pytorch.That's what I need for now,right in time.BTW，here is my answer for you question above,attention method is explained [here](https://arxiv.org/abs/1809.02983) and the code corresponding to this is \n\n\nI found this in a blog written in mainland China if you are interested [here](https://blog.csdn.net/qq_34914551/article/details/90350063)", "votes": null }, { "id": "677441", "postDate": "11/20/2019 06:47:36", "content": "Kmon,you don't really want to bring this out.BTW congrats on the GM
", "rawMarkdown": "Kmon,you don't really want to bring this out.BTW congrats on the GM", "votes": null }, { "id": "677466", "postDate": "11/20/2019 07:49:44", "content": "Thanks, I got it.
", "rawMarkdown": "Thanks, I got it.", "votes": null } ], "comments": [ { "id": 676156, "author_name": "phoenix9032", "author_url": "", "post_date": "11/19/2019 01:58:48", "content": "ouch ! you just missed one silver .I know that feeling . But for what its worth , i think you are not really a \"Overfit Queen \" .. you jumped up 31 places :)
", "votes": null, "replies": [ { "id": 676159, "author_name": "sj626591833", "author_url": "", "post_date": "11/19/2019 02:03:11", "content": "Yeah,I named myself this one cuz I got an over one thousand drop in the Lanl earthquake——my first round in this community,yet so hard to deal with time series data
", "votes": null, "replies": [] } ] }, { "id": 676815, "author_name": "raghaw", "author_url": "", "post_date": "11/19/2019 14:23:46", "content": "You were just one step ahead of me when competition ended, why your team name removed from leaderboard?\nMy Suggestion to train a model if you don't know whether it will be completed in 9 hours.\n1. Save the states of model, optimizer and scheduler after each epoch.\n2. Save the states of model, optimizer and scheduler of best model till that epoch.\n3. write your log in a plain text file in unbuffered mode, since kernel logs won't be accessible if your kernel goes dead.
\n\nif your kernel becomes dead due to time expiration you will still have above 3 files and based on logs you can decides whether you can further trained it or not. if you wants to train it for few more epoch just load the weights of last epoch of your model and states of optimizer and scheduler, and start to train it further.
", "votes": null, "replies": [ { "id": 677290, "author_name": "sj626591833", "author_url": "", "post_date": "11/20/2019 02:11:46", "content": "Firstly congrats on the silver. Also we are so appreciated for your advice,will give it a try in our next competition.
\n\nWe are not able to survive the LB cleaning,yet we do not deseve it because we shared are results with our friends in the other team who should be merged to us before the merging deadline and broke the rules I was not familiar with before.Yet we don't have a chance to do this again.
\n\nHere is our ranking before the LB cleaning
\n\nSong is a friend of us actully.I mean she is a fabulous guy.She got a great insight upon this competition.
Here are parts of our proofs for this competition.I mean we stand as a team to shed sweat,paying as much as effort as you guys,working hard trying to find the solution.
\n
\n
\n
Basically, our method was based on @gogo827jz kernel And thx him for sharing his wonderful ideas
\n\nTo make it clear,we are not that kind of cheating guys,just made a mistake which should not be made in this community.
\n\nI do not feel any pity now cuz I've been there.And glad to learn and share informations with all you guys,and I will continue being humble,using knowledge to arm myself,and sharing everything I've got with you cuz I've learned so much from this community.
", "votes": null, "replies": [] }, { "id": 677346, "author_name": "raghaw", "author_url": "", "post_date": "11/20/2019 04:03:35", "content": "Its really sad to hear this!!! Best of luck for next competitions.\nI have one question, if you can answer it will be a valuable knowledge for me.\nwhat is attention method to the decoder?
", "votes": null, "replies": [] }, { "id": 677355, "author_name": "raghaw", "author_url": "", "post_date": "11/20/2019 04:16:54", "content": "I use google colab since I don't have my own GPU. Since colab can goes off any time and therefore to prevent any loss of computation, I use follwong code to save my models, hope it will be help full to you.
\n\n
Wow,that's so nice of you.That code is so wonderful,and I'm gonna write it down to my notes which can check the best model to save according to the cv score with pytorch.That's what I need for now,right in time.BTW，here is my answer for you question above,attention method is explained here and the code corresponding to this is \n
I found this in a blog written in mainland China if you are interested here
", "votes": null, "replies": [] }, { "id": 677466, "author_name": "raghaw", "author_url": "", "post_date": "11/20/2019 07:49:44", "content": "Thanks, I got it.
", "votes": null, "replies": [] } ] }, { "id": 677335, "author_name": "naivelamb", "author_url": "", "post_date": "11/20/2019 03:55:12", "content": ">now that you choose to be a part of this comp,you should accept this unfairness
\n\nI don't get it, where is the unfairness in this competition?
", "votes": null, "replies": [ { "id": 677441, "author_name": "sj626591833", "author_url": "", "post_date": "11/20/2019 06:47:36", "content": "Kmon,you don't really want to bring this out.BTW congrats on the GM
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "676146": "It's my very first CV competition in Kaggle,APTOS blindess is so hard so I gave it up very early,but I 've been doing my best in this segmentation comp. \n##########################################\nPity? yes,I could have a chance to get a silver in the end,there was five kernels running KFold last day(UTC+8) which I think will eventually give a boost to my result,but unfortunately,three of them were dead due to the time limitation ,so I couldn't make full use of all my resources,but I got to say this is part of this competition,like Yakovlev\n saied [here](https://www.kaggle.com/c/ashrae-energy-prediction/discussion/117016#latest-674872) ,now that you choose to be a part of this comp,you should accept this unfairness,BTW,I'm a big fan of him and Chris,their insights of these competitions are so great.I learned a lot from them personally.\n##########################################\nWhat I learned from this comp is that you need to do a handful of job to get this done,I tried a lot of combinations of different encoders and decoders,and it turned out that efficientnet has a better performance, and due to the GPU time limitation .my teammate could not add attention method to the decoder and for sure this can give a boost.And it's drawing to the deadline when I tried Kfold.The reason y I can't try it earlier is the same thing due to the time limitation,so I have to use 5 kernels to save 5 models, and do ensemble later.Still got a lot of work to do in the end ,haven't found a good combination of different classifier that can boost my score. But for sure ,this can give a boost\n##########################################\nFinally,glad to win an 'expert' title in this wonderful community,you guys are just so Amazingggg!!!!", "676156": "ouch ! you just missed one silver .I know that feeling . But for what its worth , i think you are not really a \"Overfit Queen \" .. you jumped up 31 places :)", "676159": "Yeah,I named myself this one cuz I got an over one thousand drop in the Lanl earthquake——my first round in this community,yet so hard to deal with time series data", "676815": "You were just one step ahead of me when competition ended, why your team name removed from leaderboard?\nMy Suggestion to train a model if you don't know whether it will be completed in 9 hours.\n1. Save the states of model, optimizer and scheduler after each epoch.\n2. Save the states of model, optimizer and scheduler of best model till that epoch.\n3. write your log in a plain text file in unbuffered mode, since kernel logs won't be accessible if your kernel goes dead.\n\nif your kernel becomes dead due to time expiration you will still have above 3 files and based on logs you can decides whether you can further trained it or not. if you wants to train it for few more epoch just load the weights of last epoch of your model and states of optimizer and scheduler, and start to train it further.", "677290": "Firstly congrats on the silver. Also we are so appreciated for your advice,will give it a try in our next competition. \n\nWe are not able to survive the LB cleaning,yet we do not deseve it because we shared are results with our friends in the other team who should be merged to us before the merging deadline and broke the rules I was not familiar with before.Yet we don't have a chance to do this again. \n\nHere is our ranking before the LB cleaning \n\nSong is a friend of us actully.I mean she is a fabulous guy.She got a great insight upon this competition. \n\nHere are parts of our proofs for this competition.I mean we stand as a team to shed sweat,paying as much as effort as you guys,working hard trying to find the solution. \n \n \n\n\nBasically, our method was based on @gogo827jz [kernel](https://www.kaggle.com/gogo827jz/resunet-keras-with-some-new-ideas) And thx him for sharing his wonderful ideas\n\nTo make it clear,we are not that kind of cheating guys,just made a mistake which should not be made in this community. \n\nI do not feel any pity now cuz I've been there.And glad to learn and share informations with all you guys,and I will continue being humble,using knowledge to arm myself,and sharing everything I've got with you cuz I've learned so much from this community.", "677335": ">now that you choose to be a part of this comp,you should accept this unfairness\n\nI don't get it, where is the unfairness in this competition?", "677346": "Its really sad to hear this!!! Best of luck for next competitions.\nI have one question, if you can answer it will be a valuable knowledge for me.\nwhat is attention method to the decoder?", "677355": "I use google colab since I don't have my own GPU. Since colab can goes off any time and therefore to prevent any loss of computation, I use follwong code to save my models, hope it will be help full to you.\n\n\n\n", "677438": "Wow,that's so nice of you.That code is so wonderful,and I'm gonna write it down to my notes which can check the best model to save according to the cv score with pytorch.That's what I need for now,right in time.BTW，here is my answer for you question above,attention method is explained [here](https://arxiv.org/abs/1809.02983) and the code corresponding to this is \n\n\nI found this in a blog written in mainland China if you are interested [here](https://blog.csdn.net/qq_34914551/article/details/90350063)", "677441": "Kmon,you don't really want to bring this out.BTW congrats on the GM", "677466": "Thanks, I got it." }, "source": "meta" }