This was my 2nd competition in which I invested time. My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts. I did training in colab and tried to follow his recommendation of first experimenting with lower resolution images and simpler models. Also, he posted a link to previous years winners where they used ensembles of models with progressively more complex model architectures using progressively higher resolutions. That idea did not work that well for me, the lower complexity models did not perform very well. But I stuck with them for the final submission and weighted the lower models (BO with 128, B1 with 192 etc, ) very very low.
\nI do not think this gold medal is deserved. I was hoping for a top 20% finish. I think that would have been a fair reflection of my intellectual contribution. They're still working with the results, so it could still be that the final standing is more accurate. I think the one thing I did learn is that ultimately I need to be able to write notebooks like the TFR Triple Stratified from scratch. Similarly, I learned from various discussion posts threads.
", "messages": [ { "id": "974858", "postDate": "08/18/2020 04:30:28", "content": "This was my 2nd competition in which I invested time. My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts. I did training in colab and tried to follow his recommendation of first experimenting with lower resolution images and simpler models. Also, he posted a link to previous years winners where they used ensembles of models with progressively more complex model architectures using progressively higher resolutions. That idea did not work that well for me, the lower complexity models did not perform very well. But I stuck with them for the final submission and weighted the lower models (BO with 128, B1 with 192 etc, ) very very low.
\nI do not think this gold medal is deserved. I was hoping for a top 20% finish. I think that would have been a fair reflection of my intellectual contribution. They're still working with the results, so it could still be that the final standing is more accurate. I think the one thing I did learn is that ultimately I need to be able to write notebooks like the TFR Triple Stratified from scratch. Similarly, I learned from various discussion posts threads.
", "rawMarkdown": "This was my 2nd competition in which I invested time. My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts. I did training in colab and tried to follow his recommendation of first experimenting with lower resolution images and simpler models. Also, he posted a link to previous years winners where they used ensembles of models with progressively more complex model architectures using progressively higher resolutions. That idea did not work that well for me, the lower complexity models did not perform very well. But I stuck with them for the final submission and weighted the lower models (BO with 128, B1 with 192 etc, ) very very low. \n\nI do not think this gold medal is deserved. I was hoping for a top 20% finish. I think that would have been a fair reflection of my intellectual contribution. They're still working with the results, so it could still be that the final standing is more accurate. I think the one thing I did learn is that ultimately I need to be able to write notebooks like the TFR Triple Stratified from scratch. Similarly, I learned from various discussion posts threads.", "votes": null }, { "id": "974871", "postDate": "08/18/2020 04:35:14", "content": "Take what you are getting with open arms . if you are honest and learnt something that's good . A leaderboard position depends upon not only what you have done , but also what others have done. We also sometimes get unexpected results, sometimes not in our favor . Considering all these celebrate , enjoy ,learn and give back your learning to community , so that it can help someone else in some other competition .
", "rawMarkdown": "Take what you are getting with open arms . if you are honest and learnt something that's good . A leaderboard position depends upon not only what you have done , but also what others have done. We also sometimes get unexpected results, sometimes not in our favor . Considering all these celebrate , enjoy ,learn and give back your learning to community , so that it can help someone else in some other competition .", "votes": null }, { "id": "975075", "postDate": "08/18/2020 06:14:51", "content": "@sebastianji You were pretty diligent in this competition to follow all of Chris's advice and read all the discussions carefully. It would be great if you can also publish some of those scripts / notebooks on Kaggle so that others can look at your Gold medal winning solution and further learn from it.
", "rawMarkdown": "sebastianji You were pretty diligent in this competition to follow all of Chris's advice and read all the discussions carefully. It would be great if you can also publish some of those scripts / notebooks on Kaggle so that others can look at your Gold medal winning solution and further learn from it.", "votes": null }, { "id": "975129", "postDate": "08/18/2020 07:00:17", "content": "I only used Chris Deotte's Triple Stratified Kfolds with TFRecords. 5-folds for all models with consistent specifications like he suggested. To this notebook I also added his code from TFRecord Experiments - Upsample and Coarse Dropout. He also provided a link to last year's winner solution, I tried to mimic that starting from B0 with 128. not sure if this was a good idea, the models performed poorly I weighted them very low. So, in brief I made just changes to his published kernels.
\nI wanted to use the notebook on how to create 1024x1024 TFRecords , but the competition provide 1024 size TFRecords, so I did not need to create them. Here is my notebook for using 1024x1024 TFRecords . but its exactly the same.
", "rawMarkdown": "I only used Chris Deotte's [Triple Stratified Kfolds with TFRecords](https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). 5-folds for all models with consistent specifications like he suggested. To this notebook I also added his code from [TFRecord Experiments - Upsample and Coarse Dropout](https://www.kaggle.com/cdeotte/tfrecord-experiments-upsample-and-coarse-dropout). He also provided a link to last year's winner solution, I tried to mimic that starting from B0 with 128. not sure if this was a good idea, the models performed poorly I weighted them very low. So, in brief I made just changes to his published kernels. \n\nI wanted to use the notebook on how to create 1024x1024 TFRecords [](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4), but the competition provide 1024 size TFRecords, so I did not need to create them. Here is my notebook for using [1024x1024 TFRecords ](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4). but its exactly the same.", "votes": null }, { "id": "975186", "postDate": "08/18/2020 07:29:47", "content": "It is well deserved @sebastianji , because you were able to achieve this score by doing experiments from the information available publicly which others hadn't.
", "rawMarkdown": "It is well deserved @sebastianji , because you were able to achieve this score by doing experiments from the information available publicly which others hadn't.", "votes": null }, { "id": "975236", "postDate": "08/18/2020 08:00:09", "content": "good job!
\ni had a pile of notes from reading the discussion too, but iterate experiments using pytorch is tooo low.
", "rawMarkdown": "good job! \n\ni had a pile of notes from reading the discussion too, but iterate experiments using pytorch is tooo low.", "votes": null }, { "id": "975261", "postDate": "08/18/2020 08:10:12", "content": "Hi J Seb. I really applaud your modesty and honesty. I was lucky too (though not gold level :) ) and have been feeling the same way this morning.
\nI did similar to you - taking careful note of Chris's notebooks and running my own experiments of image and model size, dropout and upsampling settings. I also tried some other things like an inception and a Image Effnet concat'd with metadata but in the end these were not quite as good as the above.
\nI must admit as a newbie I got tempted by the last minute 0.96 blends that were published and submitted one to the public LB - am pleased to say that my silver is not from one of those though as I really feel I would be feeling, if not ashamed, then certainly not proud if I got a medal from that.
\nAnyway, we are here to learnt and this competition has really taught me how to use TFrecords and cross validation and why the latter is so important.
", "rawMarkdown": "Hi J Seb. I really applaud your modesty and honesty. I was lucky too (though not gold level :) ) and have been feeling the same way this morning.\n\nI did similar to you - taking careful note of Chris's notebooks and running my own experiments of image and model size, dropout and upsampling settings. I also tried some other things like an inception and a Image Effnet concat'd with metadata but in the end these were not quite as good as the above.\n\nI must admit as a newbie I got tempted by the last minute 0.96 blends that were published and submitted one to the public LB - am pleased to say that my silver is not from one of those though as I really feel I would be feeling, if not ashamed, then certainly not proud if I got a medal from that.\n\nAnyway, we are here to learnt and this competition has really taught me how to use TFrecords and cross validation and why the latter is so important.", "votes": null }, { "id": "975296", "postDate": "08/18/2020 08:28:34", "content": ":-) Standing on the shoulders of Grandmasters.
", "rawMarkdown": ":-) Standing on the shoulders of Grandmasters.", "votes": null }, { "id": "975491", "postDate": "08/18/2020 10:15:48", "content": "It is your humility speaking @sebastianji
\nIt is obvious that you have done careful groundwork and evaluated various ideas before reaching this stage so the medal is richly deserved. No doubt posts by Chris would have helped a lot, but that does not in any way diminish your efforts. If at all, you should only resolve to give back to the community the way Chris would have..
\nWould love to see a detailed analysis of your approaches, learnings and experiences. A gold medal by a 'contributor' is no mean achievement
\n\nMy strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.
\n
πππ
\nI fell from my chair while reading this.
\nNot kidding.
", "rawMarkdown": "> My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.\n\nπππ\n\nI fell from my chair while reading this. \n\nNot kidding.", "votes": null }, { "id": "975677", "postDate": "08/18/2020 12:20:39", "content": "well, he is honest :)
\nbut seriously if it were that easy, then anyone and everyone would be getting golds. I am sure a lot more must have gone into it..The posts by Chris may have been a starting point..
", "rawMarkdown": "well, he is honest :)\n\nbut seriously if it were that easy, then anyone and everyone would be getting golds. I am sure a lot more must have gone into it..The posts by Chris may have been a starting point..", "votes": null }, { "id": "975734", "postDate": "08/18/2020 12:51:50", "content": "Congratulations @sebastianji and I admire your humility. You stuck to your method and you were disciplined not trying to chase the Public LB. That is a strong personality attribute which will serve you well as you advance in data science. This result will also encourage you along that journey. Well done!
", "rawMarkdown": "Congratulations @sebastianji and I admire your humility. You stuck to your method and you were disciplined not trying to chase the Public LB. That is a strong personality attribute which will serve you well as you advance in data science. This result will also encourage you along that journey. Well done!", "votes": null }, { "id": "975759", "postDate": "08/18/2020 13:06:44", "content": "Thanks, I think it speaks to the strength of the TFR notebook. I did follow recommendations in those notebooks and discussion posts associated with them such as experimenting with simpler models and lower resolutions . Regarding approach, I tried to implement a link to a post in the TFR notebook about last year's winner's strategy. ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3704899%2Fd7b303ee4407cbcb63711df74200dc86%2Flast_year.png?generation=1597754631529458&alt=media\")
so i tried to stack something similar B0-128, B0-192, B2-192, B1-256 etc. But the models for last years winners solution did not differ by much, whereas here found that there were differences among models by as much as 5%. So I included them in the final solution but with really low weights in the ensemble.
Then I also included some models that I had tried out initially when trying to implement the TFR notebook but had good performance. In all these models I implemented coarse dropout and upsampling of positive images from Chris notebook, except for 1024x1024 B7 due to TPU time constraints on Kaggle.
\nIn the end I had 15 models and posted a question in a discussion thread about how to ensemble them. I got really good input but I had not saved my OOF prediction files, so thats something I learnt. Without a better way to ensemble models, I ordered my models by performance and up-voted higher performing models.
", "rawMarkdown": "Thanks, I think it speaks to the strength of the TFR notebook. I did follow recommendations in those notebooks and discussion posts associated with them such as experimenting with simpler models and lower resolutions . Regarding approach, I tried to implement a link to a post in the TFR notebook about last year's winner's strategy.  so i tried to stack something similar B0-128, B0-192, B2-192, B1-256 etc. But the models for last years winners solution did not differ by much, whereas here found that there were differences among models by as much as 5%. So I included them in the final solution but with really low weights in the ensemble. \n\nThen I also included some models that I had tried out initially when trying to implement the TFR notebook but had good performance. In all these models I implemented coarse dropout and upsampling of positive images from Chris notebook, except for [1024x1024 B7](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4) due to TPU time constraints on Kaggle. \n\nIn the end I had 15 models and posted a question in a [discussion thread about how to ensemble them](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/174058). I got really good input but I had not saved my OOF prediction files, so thats something I learnt. Without a better way to ensemble models, I ordered my models by performance and up-voted higher performing models.", "votes": null }, { "id": "975991", "postDate": "08/18/2020 15:09:26", "content": "WOW, I just checked your profile. It's great to see a MIZZOU buddy on Kaggle. Congrats!
", "rawMarkdown": "WOW, I just checked your profile. It's great to see a MIZZOU buddy on Kaggle. Congrats!", "votes": null }, { "id": "976002", "postDate": "08/18/2020 15:17:00", "content": "Thanks much!!
", "rawMarkdown": "Thanks much!!", "votes": null }, { "id": "976747", "postDate": "08/19/2020 04:28:55", "content": "Awesome! I see you ended up in Top 10. Congrats once again and all the very best for your future competitions
", "rawMarkdown": "Awesome! I see you ended up in Top 10. Congrats once again and all the very best for your future competitions", "votes": null }, { "id": "980856", "postDate": "08/21/2020 23:56:26", "content": "\n\nMy strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.
\n
Great strategy!! πYou did my strategy better than I did πCongrats on your awesome solo gold finish.
", "rawMarkdown": "> My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.\n\nGreat strategy!! πYou did my strategy better than I did πCongrats on your awesome solo gold finish.", "votes": null }, { "id": "980968", "postDate": "08/22/2020 04:12:52", "content": "ππ My final placement seems written for a comedy bit. Thanks so much again for sharing your awesome work and insights. Very indebted.
", "rawMarkdown": "ππ My final placement seems written for a comedy bit. Thanks so much again for sharing your awesome work and insights. Very indebted.", "votes": null }, { "id": "981142", "postDate": "08/22/2020 07:49:40", "content": "Congratulations! you ended up in top 10 :].
", "rawMarkdown": "Congratulations! you ended up in top 10 :].", "votes": null }, { "id": "982142", "postDate": "08/23/2020 05:15:25", "content": "Congrats Mate!
", "rawMarkdown": "Congrats Mate!", "votes": null }, { "id": "982541", "postDate": "08/23/2020 13:01:45", "content": "Thank you!!
", "rawMarkdown": "Thank you!!", "votes": null } ], "comments": [ { "id": 974871, "author_name": "phoenix9032", "author_url": "", "post_date": "08/18/2020 04:35:14", "content": "Take what you are getting with open arms . if you are honest and learnt something that's good . A leaderboard position depends upon not only what you have done , but also what others have done. We also sometimes get unexpected results, sometimes not in our favor . Considering all these celebrate , enjoy ,learn and give back your learning to community , so that it can help someone else in some other competition .
", "votes": null, "replies": [] }, { "id": 975075, "author_name": "priteshshrivastava", "author_url": "", "post_date": "08/18/2020 06:14:51", "content": "@sebastianji You were pretty diligent in this competition to follow all of Chris's advice and read all the discussions carefully. It would be great if you can also publish some of those scripts / notebooks on Kaggle so that others can look at your Gold medal winning solution and further learn from it.
", "votes": null, "replies": [ { "id": 975129, "author_name": "sebastianji", "author_url": "", "post_date": "08/18/2020 07:00:17", "content": "I only used Chris Deotte's Triple Stratified Kfolds with TFRecords. 5-folds for all models with consistent specifications like he suggested. To this notebook I also added his code from TFRecord Experiments - Upsample and Coarse Dropout. He also provided a link to last year's winner solution, I tried to mimic that starting from B0 with 128. not sure if this was a good idea, the models performed poorly I weighted them very low. So, in brief I made just changes to his published kernels.
\nI wanted to use the notebook on how to create 1024x1024 TFRecords , but the competition provide 1024 size TFRecords, so I did not need to create them. Here is my notebook for using 1024x1024 TFRecords . but its exactly the same.
", "votes": null, "replies": [] } ] }, { "id": 975186, "author_name": "karrak3256", "author_url": "", "post_date": "08/18/2020 07:29:47", "content": "It is well deserved @sebastianji , because you were able to achieve this score by doing experiments from the information available publicly which others hadn't.
", "votes": null, "replies": [] }, { "id": 975236, "author_name": "yimacs", "author_url": "", "post_date": "08/18/2020 08:00:09", "content": "good job!
\ni had a pile of notes from reading the discussion too, but iterate experiments using pytorch is tooo low.
", "votes": null, "replies": [] }, { "id": 975261, "author_name": "cascadenite", "author_url": "", "post_date": "08/18/2020 08:10:12", "content": "Hi J Seb. I really applaud your modesty and honesty. I was lucky too (though not gold level :) ) and have been feeling the same way this morning.
\nI did similar to you - taking careful note of Chris's notebooks and running my own experiments of image and model size, dropout and upsampling settings. I also tried some other things like an inception and a Image Effnet concat'd with metadata but in the end these were not quite as good as the above.
\nI must admit as a newbie I got tempted by the last minute 0.96 blends that were published and submitted one to the public LB - am pleased to say that my silver is not from one of those though as I really feel I would be feeling, if not ashamed, then certainly not proud if I got a medal from that.
\nAnyway, we are here to learnt and this competition has really taught me how to use TFrecords and cross validation and why the latter is so important.
", "votes": null, "replies": [ { "id": 975296, "author_name": "sebastianji", "author_url": "", "post_date": "08/18/2020 08:28:34", "content": ":-) Standing on the shoulders of Grandmasters.
", "votes": null, "replies": [] } ] }, { "id": 975491, "author_name": "allohvk", "author_url": "", "post_date": "08/18/2020 10:15:48", "content": "It is your humility speaking @sebastianji
\nIt is obvious that you have done careful groundwork and evaluated various ideas before reaching this stage so the medal is richly deserved. No doubt posts by Chris would have helped a lot, but that does not in any way diminish your efforts. If at all, you should only resolve to give back to the community the way Chris would have..
\nWould love to see a detailed analysis of your approaches, learnings and experiences. A gold medal by a 'contributor' is no mean achievement
Thanks, I think it speaks to the strength of the TFR notebook. I did follow recommendations in those notebooks and discussion posts associated with them such as experimenting with simpler models and lower resolutions . Regarding approach, I tried to implement a link to a post in the TFR notebook about last year's winner's strategy. so i tried to stack something similar B0-128, B0-192, B2-192, B1-256 etc. But the models for last years winners solution did not differ by much, whereas here found that there were differences among models by as much as 5%. So I included them in the final solution but with really low weights in the ensemble.
Then I also included some models that I had tried out initially when trying to implement the TFR notebook but had good performance. In all these models I implemented coarse dropout and upsampling of positive images from Chris notebook, except for 1024x1024 B7 due to TPU time constraints on Kaggle.
\nIn the end I had 15 models and posted a question in a discussion thread about how to ensemble them. I got really good input but I had not saved my OOF prediction files, so thats something I learnt. Without a better way to ensemble models, I ordered my models by performance and up-voted higher performing models.
", "votes": null, "replies": [] }, { "id": 976747, "author_name": "allohvk", "author_url": "", "post_date": "08/19/2020 04:28:55", "content": "Awesome! I see you ended up in Top 10. Congrats once again and all the very best for your future competitions
", "votes": null, "replies": [] } ] }, { "id": 975623, "author_name": "rohitagarwal", "author_url": "", "post_date": "08/18/2020 11:39:47", "content": "\n\nMy strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.
\n
πππ
\nI fell from my chair while reading this.
\nNot kidding.
", "votes": null, "replies": [ { "id": 975677, "author_name": "allohvk", "author_url": "", "post_date": "08/18/2020 12:20:39", "content": "well, he is honest :)
\nbut seriously if it were that easy, then anyone and everyone would be getting golds. I am sure a lot more must have gone into it..The posts by Chris may have been a starting point..
", "votes": null, "replies": [] } ] }, { "id": 975734, "author_name": "jsyphil", "author_url": "", "post_date": "08/18/2020 12:51:50", "content": "Congratulations @sebastianji and I admire your humility. You stuck to your method and you were disciplined not trying to chase the Public LB. That is a strong personality attribute which will serve you well as you advance in data science. This result will also encourage you along that journey. Well done!
", "votes": null, "replies": [] }, { "id": 975991, "author_name": "waylongo", "author_url": "", "post_date": "08/18/2020 15:09:26", "content": "WOW, I just checked your profile. It's great to see a MIZZOU buddy on Kaggle. Congrats!
", "votes": null, "replies": [ { "id": 976002, "author_name": "sebastianji", "author_url": "", "post_date": "08/18/2020 15:17:00", "content": "Thanks much!!
", "votes": null, "replies": [] } ] }, { "id": 980856, "author_name": "cdeotte", "author_url": "", "post_date": "08/21/2020 23:56:26", "content": "\n\nMy strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.
\n
Great strategy!! πYou did my strategy better than I did πCongrats on your awesome solo gold finish.
", "votes": null, "replies": [ { "id": 980968, "author_name": "sebastianji", "author_url": "", "post_date": "08/22/2020 04:12:52", "content": "ππ My final placement seems written for a comedy bit. Thanks so much again for sharing your awesome work and insights. Very indebted.
", "votes": null, "replies": [] } ] }, { "id": 981142, "author_name": "rifat963", "author_url": "", "post_date": "08/22/2020 07:49:40", "content": "Congratulations! you ended up in top 10 :].
", "votes": null, "replies": [] }, { "id": 982142, "author_name": "raoofnaushad", "author_url": "", "post_date": "08/23/2020 05:15:25", "content": "Congrats Mate!
", "votes": null, "replies": [ { "id": 982541, "author_name": "sebastianji", "author_url": "", "post_date": "08/23/2020 13:01:45", "content": "Thank you!!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "974858": "This was my 2nd competition in which I invested time. My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts. I did training in colab and tried to follow his recommendation of first experimenting with lower resolution images and simpler models. Also, he posted a link to previous years winners where they used ensembles of models with progressively more complex model architectures using progressively higher resolutions. That idea did not work that well for me, the lower complexity models did not perform very well. But I stuck with them for the final submission and weighted the lower models (BO with 128, B1 with 192 etc, ) very very low. \n\nI do not think this gold medal is deserved. I was hoping for a top 20% finish. I think that would have been a fair reflection of my intellectual contribution. They're still working with the results, so it could still be that the final standing is more accurate. I think the one thing I did learn is that ultimately I need to be able to write notebooks like the TFR Triple Stratified from scratch. Similarly, I learned from various discussion posts threads.", "974871": "Take what you are getting with open arms . if you are honest and learnt something that's good . A leaderboard position depends upon not only what you have done , but also what others have done. We also sometimes get unexpected results, sometimes not in our favor . Considering all these celebrate , enjoy ,learn and give back your learning to community , so that it can help someone else in some other competition .", "975075": "sebastianji You were pretty diligent in this competition to follow all of Chris's advice and read all the discussions carefully. It would be great if you can also publish some of those scripts / notebooks on Kaggle so that others can look at your Gold medal winning solution and further learn from it.", "975129": "I only used Chris Deotte's [Triple Stratified Kfolds with TFRecords](https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). 5-folds for all models with consistent specifications like he suggested. To this notebook I also added his code from [TFRecord Experiments - Upsample and Coarse Dropout](https://www.kaggle.com/cdeotte/tfrecord-experiments-upsample-and-coarse-dropout). He also provided a link to last year's winner solution, I tried to mimic that starting from B0 with 128. not sure if this was a good idea, the models performed poorly I weighted them very low. So, in brief I made just changes to his published kernels. \n\nI wanted to use the notebook on how to create 1024x1024 TFRecords [](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4), but the competition provide 1024 size TFRecords, so I did not need to create them. Here is my notebook for using [1024x1024 TFRecords ](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4). but its exactly the same.", "975186": "It is well deserved @sebastianji , because you were able to achieve this score by doing experiments from the information available publicly which others hadn't.", "975236": "good job! \n\ni had a pile of notes from reading the discussion too, but iterate experiments using pytorch is tooo low.", "975261": "Hi J Seb. I really applaud your modesty and honesty. I was lucky too (though not gold level :) ) and have been feeling the same way this morning.\n\nI did similar to you - taking careful note of Chris's notebooks and running my own experiments of image and model size, dropout and upsampling settings. I also tried some other things like an inception and a Image Effnet concat'd with metadata but in the end these were not quite as good as the above.\n\nI must admit as a newbie I got tempted by the last minute 0.96 blends that were published and submitted one to the public LB - am pleased to say that my silver is not from one of those though as I really feel I would be feeling, if not ashamed, then certainly not proud if I got a medal from that.\n\nAnyway, we are here to learnt and this competition has really taught me how to use TFrecords and cross validation and why the latter is so important.", "975296": ":-) Standing on the shoulders of Grandmasters.", "975491": "It is your humility speaking @sebastianji \nIt is obvious that you have done careful groundwork and evaluated various ideas before reaching this stage so the medal is richly deserved. No doubt posts by Chris would have helped a lot, but that does not in any way diminish your efforts. If at all, you should only resolve to give back to the community the way Chris would have..\nWould love to see a detailed analysis of your approaches, learnings and experiences. A gold medal by a 'contributor' is no mean achievement", "975623": "> My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.\n\nπππ\n\nI fell from my chair while reading this. \n\nNot kidding.", "975677": "well, he is honest :)\n\nbut seriously if it were that easy, then anyone and everyone would be getting golds. I am sure a lot more must have gone into it..The posts by Chris may have been a starting point..", "975734": "Congratulations @sebastianji and I admire your humility. You stuck to your method and you were disciplined not trying to chase the Public LB. That is a strong personality attribute which will serve you well as you advance in data science. This result will also encourage you along that journey. Well done!", "975759": "Thanks, I think it speaks to the strength of the TFR notebook. I did follow recommendations in those notebooks and discussion posts associated with them such as experimenting with simpler models and lower resolutions . Regarding approach, I tried to implement a link to a post in the TFR notebook about last year's winner's strategy.  so i tried to stack something similar B0-128, B0-192, B2-192, B1-256 etc. But the models for last years winners solution did not differ by much, whereas here found that there were differences among models by as much as 5%. So I included them in the final solution but with really low weights in the ensemble. \n\nThen I also included some models that I had tried out initially when trying to implement the TFR notebook but had good performance. In all these models I implemented coarse dropout and upsampling of positive images from Chris notebook, except for [1024x1024 B7](https://www.kaggle.com/sebastianji/triple-stratified-kfold-js-1024-ef7-fold4) due to TPU time constraints on Kaggle. \n\nIn the end I had 15 models and posted a question in a [discussion thread about how to ensemble them](https://www.kaggle.com/c/siim-isic-melanoma-classification/discussion/174058). I got really good input but I had not saved my OOF prediction files, so thats something I learnt. Without a better way to ensemble models, I ordered my models by performance and up-voted higher performing models.", "975991": "WOW, I just checked your profile. It's great to see a MIZZOU buddy on Kaggle. Congrats!", "976002": "Thanks much!!", "976747": "Awesome! I see you ended up in Top 10. Congrats once again and all the very best for your future competitions", "980856": "> My strategy was this: I ordered notebooks and discussion posts by Chris Deotte @cdeotte by date and basically implemented each of those kernels and suggestions from discussion posts.\n\nGreat strategy!! πYou did my strategy better than I did πCongrats on your awesome solo gold finish.", "980968": "ππ My final placement seems written for a comedy bit. Thanks so much again for sharing your awesome work and insights. Very indebted.", "981142": "Congratulations! you ended up in top 10 :].", "982142": "Congrats Mate!", "982541": "Thank you!!" }, "source": "meta" }