Since it seems that I won't end-up at medal range this time, I would like to post this before we get the rain of awesome reports after the results.
\nThis competition was very special to me, I joined because after the melanoma competition I felt that I had a little more to explore in the computer vision field, and it is basically what I have don't at the first two months, just trying cool things (GANs, SCL, multi-tasking), after the new year I got less time to put here, and I was not able to really compete. But this was probably the competition where I shared most notebooks and even got 4 gold medals (most of my previous kernels stopped a little short from gold 😓), now it even seems possible to become GM, so I would like to thank you all! 🙌👍 By looking at the comments at my kernels/discussions and the notebook count using my datasets it seems that I was able to support many people during this competition, this really feels great!
As usual, I made a GitHub repository with all my experiments, scripts, and many useful references, it may be interesting to give it a quick look.
\nps. I hope I did not make any mistakes creating those datasets, and messed with other people's experiments, I used them for all of mine.
\nedit [1]: One thing that I forgot to mention, during this competition I became a Google developer expert on ML, it was not a direct result of this competition, but my work on Kaggle had a great impact on that, no I am more able to support the community also outside of Kaggle!
", "messages": [ { "id": 1209438, "postDate": "2021-02-18T22:32:44.363Z", "content": "Since it seems that I won't end-up at medal range this time, I would like to post this before we get the rain of awesome reports after the results.
\nThis competition was very special to me, I joined because after the melanoma competition I felt that I had a little more to explore in the computer vision field, and it is basically what I have don't at the first two months, just trying cool things (GANs, SCL, multi-tasking), after the new year I got less time to put here, and I was not able to really compete. But this was probably the competition where I shared most notebooks and even got 4 gold medals (most of my previous kernels stopped a little short from gold 😓), now it even seems possible to become GM, so I would like to thank you all! 🙌👍 By looking at the comments at my kernels/discussions and the notebook count using my datasets it seems that I was able to support many people during this competition, this really feels great!
As usual, I made a GitHub repository with all my experiments, scripts, and many useful references, it may be interesting to give it a quick look.
\nps. I hope I did not make any mistakes creating those datasets, and messed with other people's experiments, I used them for all of mine.
\nedit [1]: One thing that I forgot to mention, during this competition I became a Google developer expert on ML, it was not a direct result of this competition, but my work on Kaggle had a great impact on that, no I am more able to support the community also outside of Kaggle!
", "rawMarkdown": "Since it seems that I won't end-up at medal range this time, I would like to post this before we get the rain of awesome reports after the results.\n\nThis competition was very special to me, I joined because after the melanoma competition I felt that I had a little more to explore in the computer vision field, and it is basically what I have don't at the first two months, just trying cool things `(GANs, SCL, multi-tasking)`, after the new year I got less time to put here, and I was not able to really compete. But this was probably the competition where I shared most notebooks and even got 4 gold medals (most of my previous kernels stopped a little short from gold 😓), now it even seems possible to become GM, **so I would like to thank you all! 🙌👍** By looking at the comments at my kernels/discussions and the notebook count using my datasets it seems that I was able to support many people during this competition, **this really feels great!**\n\nAs usual, I made a [GitHub repository](https://github.com/dimitreOliveira/Cassava-Leaf-Disease-Classification) with all my experiments, scripts, and many useful references, it may be interesting to give it a quick look.\n\nps. I hope I did not make any mistakes creating those datasets, and messed with other people's experiments, I used them for all of mine.\n\nedit [1]: One thing that I forgot to mention, during this competition I became a [Google developer expert on ML](https://developers.google.com/community/experts/directory/profile/profile-dimitre_oliveira), it was not a direct result of this competition, but my work on Kaggle had a great impact on that, no I am more able to support the community also outside of Kaggle!", "votes": 25 }, { "id": 1209444, "postDate": "2021-02-18T22:44:29.610Z", "content": "Thanks Dimitre for sharing great notebooks in this comp and Rainforest Comp. I believe you should win a $1000 TPU Star Award for your great contributions in Cassava Comp. Go ahead and submit the form here. I believe you need to submit it yourself. If i am mistaken and you need someone to nominate you, let me know and I will nominate you.
", "rawMarkdown": "Thanks Dimitre for sharing great notebooks in this comp and Rainforest Comp. I believe you should win a $1000 TPU Star Award for your great contributions in Cassava Comp. Go ahead and submit the form [here][1]. I believe you need to submit it yourself. If i am mistaken and you need someone to nominate you, let me know and I will nominate you.\n\n[1]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220244", "votes": 10, "replies": [ { "id": 1209475, "postDate": "2021-02-18T23:18:01.410Z", "content": "Thank you @cdeotte , you are an awesome guy, I don't think there is a need for the nomination, we just need to submit the form, and I will do it for sure, but anyway thanks for being available, for me, it is an honor.
\nI kinda missed your public notebooks on those competitions because you always share amazing contents, but I am always mind blown by how some people here manage to compete at multiple competitions, end up at the top places and keep a full-time work 😄, I am sure it might be a tight schedule.
", "rawMarkdown": "Thank you @cdeotte , you are an awesome guy, I don't think there is a need for the nomination, we just need to submit the form, and I will do it for sure, but anyway thanks for being available, for me, it is an honor.\n\nI kinda missed your public notebooks on those competitions because you always share amazing contents, but I am always mind blown by how some people here manage to compete at multiple competitions, end up at the top places and keep a full-time work 😄, I am sure it might be a tight schedule.", "votes": 2 } ] }, { "id": 1209453, "postDate": "2021-02-18T22:52:17.413Z", "content": "Thanks for everything you shared in this competition, I feel really sad that you weren't able to break through the top of the LB. I used your datasets and chunks of your code in my pipeline: you definitely deserved all these medals and for sure you need to compete for the TPU star prize!
", "rawMarkdown": "Thanks for everything you shared in this competition, I feel really sad that you weren't able to break through the top of the LB. I used your datasets and chunks of your code in my pipeline: you definitely deserved all these medals and for sure you need to compete for the TPU star prize!", "votes": 4, "replies": [ { "id": 1209478, "postDate": "2021-02-18T23:21:56.677Z", "content": "You're welcome @mviola , but don't feel sad this was a great learning experience, for me this is the cake, medals are just the cherry at the top, next time we will do better!
", "rawMarkdown": "You're welcome @mviola , but don't feel sad this was a great learning experience, for me this is the cake, medals are just the cherry at the top, next time we will do better!", "votes": 1 } ] }, { "id": 1211135, "postDate": "2021-02-20T01:37:19.400Z", "content": "Yep thanks hope you receive the award, I found the notebooks very helpful.
\nAlthough I was making final submissions with pytorch, actually now I can see private scores, one of the keras models was actually very strong, it seems.
\nSingle model comparison (not architecture, no 5-fold - just one set of weights)
\nKeras single model (EnetB3, based on your notebooks, dropped a few of the dodgy 'healthy' samples, added CutMix from @cdeotte notebook in flower comp, pretrained on 2019, 2 center crops TTA with slightly different zooms)
\nPriv / Public 0.8961 0.8889
Pytorch best single model - VIT
\nPriv / Public 0.8957 0.8987
Unfortunately although the CV with the Keras models was quite strong, and also the private as we can see, for whatever reason the keras scores were so much lower on the public, and I landed up being discouraged.
", "rawMarkdown": "Yep thanks hope you receive the award, I found the notebooks very helpful.\n\nAlthough I was making final submissions with pytorch, actually now I can see private scores, one of the keras models was actually very strong, it seems.\n\nSingle model comparison (not architecture, no 5-fold - just one set of weights)\n\nKeras single model (EnetB3, based on your notebooks, dropped a few of the dodgy 'healthy' samples, added CutMix from @cdeotte notebook in flower comp, pretrained on 2019, 2 center crops TTA with slightly different zooms)\nPriv / Public 0.8961 0.8889\n\nPytorch best single model - VIT\nPriv / Public 0.8957 0.8987\n\nUnfortunately although the CV with the Keras models was quite strong, and also the private as we can see, for whatever reason the keras scores were so much lower on the public, and I landed up being discouraged.", "votes": 1, "replies": [ { "id": 1211804, "postDate": "2021-02-20T15:17:20.400Z", "content": "Thanks @davidedwards1 ,
\nYeah, there is always this debate between TF and Pytorch, fortunately, the gap was very low here, but being able to code with both usually is an advantage, but let's say I am an \"only Tensorflow\" guy, thanks for your report.
", "rawMarkdown": "Thanks @davidedwards1 ,\n\nYeah, there is always this debate between TF and Pytorch, fortunately, the gap was very low here, but being able to code with both usually is an advantage, but let's say I am an \"only Tensorflow\" guy, thanks for your report." } ] }, { "id": 1210407, "postDate": "2021-02-19T12:10:16.567Z", "content": "Thanks for the great notebook!
\nMy submission was created using your notebook as a reference. My submissions are based on your notebooks, and from your notebooks I learned how to use TPU.
\nI sincerely hope you win the TPU Star Award!
That is great @kkmax1015 , I am glad I could help you.
", "rawMarkdown": "That is great @kkmax1015 , I am glad I could help you.", "votes": 1 } ] }, { "id": 1209470, "postDate": "2021-02-18T23:13:28.237Z", "content": "@dimitreoliveira Thanks for your notebook, really helped me in understanding how to do custom training in TF.
", "rawMarkdown": "@dimitreoliveira Thanks for your notebook, really helped me in understanding how to do custom training in TF.", "votes": 1, "replies": [ { "id": 1209486, "postDate": "2021-02-18T23:27:38.187Z", "content": "You're welcome @vickygoyal , this is great, after a couple of years I felt that I was really starting to get comfortable with TF when I began to work with custom training.
", "rawMarkdown": "You're welcome @vickygoyal , this is great, after a couple of years I felt that I was really starting to get comfortable with TF when I began to work with custom training.", "votes": 2 } ] }, { "id": 1209467, "postDate": "2021-02-18T23:11:25.590Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1209459, "postDate": "2021-02-18T22:59:20.003Z", "content": "Thanks for everything !
", "rawMarkdown": "Thanks for everything !", "votes": 1 } ], "comments": [ { "id": 1209444, "author_name": "Chris Deotte", "author_url": "", "post_date": "2021-02-18T22:44:29.610000", "content": "Thanks Dimitre for sharing great notebooks in this comp and Rainforest Comp. I believe you should win a $1000 TPU Star Award for your great contributions in Cassava Comp. Go ahead and submit the form here. I believe you need to submit it yourself. If i am mistaken and you need someone to nominate you, let me know and I will nominate you.
", "votes": 10, "replies": [ { "id": 1209475, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2021-02-18T23:18:01.410000", "content": "Thank you @cdeotte , you are an awesome guy, I don't think there is a need for the nomination, we just need to submit the form, and I will do it for sure, but anyway thanks for being available, for me, it is an honor.
\nI kinda missed your public notebooks on those competitions because you always share amazing contents, but I am always mind blown by how some people here manage to compete at multiple competitions, end up at the top places and keep a full-time work 😄, I am sure it might be a tight schedule.
", "votes": 2, "replies": [] } ] }, { "id": 1209453, "author_name": "Massimiliano Viola", "author_url": "", "post_date": "2021-02-18T22:52:17.413000", "content": "Thanks for everything you shared in this competition, I feel really sad that you weren't able to break through the top of the LB. I used your datasets and chunks of your code in my pipeline: you definitely deserved all these medals and for sure you need to compete for the TPU star prize!
", "votes": 4, "replies": [ { "id": 1209478, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2021-02-18T23:21:56.677000", "content": "You're welcome @mviola , but don't feel sad this was a great learning experience, for me this is the cake, medals are just the cherry at the top, next time we will do better!
", "votes": 1, "replies": [] } ] }, { "id": 1211135, "author_name": "Dave E", "author_url": "", "post_date": "2021-02-20T01:37:19.400000", "content": "Yep thanks hope you receive the award, I found the notebooks very helpful.
\nAlthough I was making final submissions with pytorch, actually now I can see private scores, one of the keras models was actually very strong, it seems.
\nSingle model comparison (not architecture, no 5-fold - just one set of weights)
\nKeras single model (EnetB3, based on your notebooks, dropped a few of the dodgy 'healthy' samples, added CutMix from @cdeotte notebook in flower comp, pretrained on 2019, 2 center crops TTA with slightly different zooms)
\nPriv / Public 0.8961 0.8889
Pytorch best single model - VIT
\nPriv / Public 0.8957 0.8987
Unfortunately although the CV with the Keras models was quite strong, and also the private as we can see, for whatever reason the keras scores were so much lower on the public, and I landed up being discouraged.
", "votes": 1, "replies": [ { "id": 1211804, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2021-02-20T15:17:20.400000", "content": "Thanks @davidedwards1 ,
\nYeah, there is always this debate between TF and Pytorch, fortunately, the gap was very low here, but being able to code with both usually is an advantage, but let's say I am an \"only Tensorflow\" guy, thanks for your report.
", "votes": 0, "replies": [] } ] }, { "id": 1210407, "author_name": "K.Kamata", "author_url": "", "post_date": "2021-02-19T12:10:16.567000", "content": "Thanks for the great notebook!
\nMy submission was created using your notebook as a reference. My submissions are based on your notebooks, and from your notebooks I learned how to use TPU.
\nI sincerely hope you win the TPU Star Award!
That is great @kkmax1015 , I am glad I could help you.
", "votes": 1, "replies": [] } ] }, { "id": 1209470, "author_name": "Vicky Goyal", "author_url": "", "post_date": "2021-02-18T23:13:28.237000", "content": "@dimitreoliveira Thanks for your notebook, really helped me in understanding how to do custom training in TF.
", "votes": 1, "replies": [ { "id": 1209486, "author_name": "DimitreOliveira", "author_url": "", "post_date": "2021-02-18T23:27:38.187000", "content": "You're welcome @vickygoyal , this is great, after a couple of years I felt that I was really starting to get comfortable with TF when I began to work with custom training.
", "votes": 2, "replies": [] } ] }, { "id": 1209467, "author_name": "", "author_url": "", "post_date": "2021-02-18T23:11:25.590000", "content": "", "votes": 1, "replies": [] }, { "id": 1209459, "author_name": "Manuel Campos", "author_url": "", "post_date": "2021-02-18T22:59:20.003000", "content": "Thanks for everything !
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "1209438": "Since it seems that I won't end-up at medal range this time, I would like to post this before we get the rain of awesome reports after the results.\n\nThis competition was very special to me, I joined because after the melanoma competition I felt that I had a little more to explore in the computer vision field, and it is basically what I have don't at the first two months, just trying cool things `(GANs, SCL, multi-tasking)`, after the new year I got less time to put here, and I was not able to really compete. But this was probably the competition where I shared most notebooks and even got 4 gold medals (most of my previous kernels stopped a little short from gold 😓), now it even seems possible to become GM, **so I would like to thank you all! 🙌👍** By looking at the comments at my kernels/discussions and the notebook count using my datasets it seems that I was able to support many people during this competition, **this really feels great!**\n\nAs usual, I made a [GitHub repository](https://github.com/dimitreOliveira/Cassava-Leaf-Disease-Classification) with all my experiments, scripts, and many useful references, it may be interesting to give it a quick look.\n\nps. I hope I did not make any mistakes creating those datasets, and messed with other people's experiments, I used them for all of mine.\n\nedit [1]: One thing that I forgot to mention, during this competition I became a [Google developer expert on ML](https://developers.google.com/community/experts/directory/profile/profile-dimitre_oliveira), it was not a direct result of this competition, but my work on Kaggle had a great impact on that, no I am more able to support the community also outside of Kaggle!", "1209444": "Thanks Dimitre for sharing great notebooks in this comp and Rainforest Comp. I believe you should win a $1000 TPU Star Award for your great contributions in Cassava Comp. Go ahead and submit the form [here][1]. I believe you need to submit it yourself. If i am mistaken and you need someone to nominate you, let me know and I will nominate you.\n\n[1]: https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/220244", "1209453": "Thanks for everything you shared in this competition, I feel really sad that you weren't able to break through the top of the LB. I used your datasets and chunks of your code in my pipeline: you definitely deserved all these medals and for sure you need to compete for the TPU star prize!", "1211135": "Yep thanks hope you receive the award, I found the notebooks very helpful.\n\nAlthough I was making final submissions with pytorch, actually now I can see private scores, one of the keras models was actually very strong, it seems.\n\nSingle model comparison (not architecture, no 5-fold - just one set of weights)\n\nKeras single model (EnetB3, based on your notebooks, dropped a few of the dodgy 'healthy' samples, added CutMix from @cdeotte notebook in flower comp, pretrained on 2019, 2 center crops TTA with slightly different zooms)\nPriv / Public 0.8961 0.8889\n\nPytorch best single model - VIT\nPriv / Public 0.8957 0.8987\n\nUnfortunately although the CV with the Keras models was quite strong, and also the private as we can see, for whatever reason the keras scores were so much lower on the public, and I landed up being discouraged.", "1210407": "Thanks for the great notebook!\nMy submission was created using your notebook as a reference. My submissions are based on your notebooks, and from your notebooks I learned how to use TPU.\nI sincerely hope you win the TPU Star Award!", "1209470": "@dimitreoliveira Thanks for your notebook, really helped me in understanding how to do custom training in TF.", "1209467": "", "1209459": "Thanks for everything !" } }