This is just a post to express my thoughts on the competition, and what I learned(this is the first time i dedicated myself to the competition).
\n\nI'll just get my approach out of the way:\nMy models alone got a 0.915 private / 0.815 public.\nI originally started this competition by copying @drhabib 's b5 kernel and toying with the thresholds and hyperparams(this got me to 0.798 public lb). Again, I was pretty new to ml/cv so I didn't know how to set up everything.\nI then started on my own approach. Guided by many great kernels from @abhishek and @ratthachat , i slowly began to write my own pipeline and training scripts. \nMy final approach actually became to be training on 2015+2019 with cropping black and using 2019 as validation(probably not the best practice haha). I then trained 3 models each for EfficientNetB2-5 where i checkpointed the model with best loss, best kappa, and the last model. I wouldve finetuned more on 2019 but kaggles kernel restrictions hit me pretty hard(no local machine :/).\nMy final ensemble was using the b3(best kappa+loss) and b4/5(all) so 8 models in total.
\n\nNow my thoughts(not technical)\n1. Always read through public kernels/discussions. These are so valuable and provide many insights and can alone probably score a bronze medal. Look through high scoring kernels and see what they do similarly and different and question why. So many people share there training/validation strategies in discussions so it is easy to modify stuff from there.\n2. Try to team up! Not only do teammates raise your score(brought me from bronze -> silver), they can teach so much. I was pretty lucky to team up with @taindow and @rishabhiitbhu (whom i cant thank enough!) because they taught me so much about machine learning, from using k-fold, different training strategies, how to analyze confusion matrices/model statistics, to how to keep track of models and their metrics.
\n\nAnyways, i hope this is helpful to new kagglers and I wish everyone luck in their kaggle adventures!
", "messages": [ { "id": "623575", "postDate": "09/11/2019 05:06:18", "content": "This is just a post to express my thoughts on the competition, and what I learned(this is the first time i dedicated myself to the competition).
\n\nI'll just get my approach out of the way:\nMy models alone got a 0.915 private / 0.815 public.\nI originally started this competition by copying @drhabib 's b5 kernel and toying with the thresholds and hyperparams(this got me to 0.798 public lb). Again, I was pretty new to ml/cv so I didn't know how to set up everything.\nI then started on my own approach. Guided by many great kernels from @abhishek and @ratthachat , i slowly began to write my own pipeline and training scripts. \nMy final approach actually became to be training on 2015+2019 with cropping black and using 2019 as validation(probably not the best practice haha). I then trained 3 models each for EfficientNetB2-5 where i checkpointed the model with best loss, best kappa, and the last model. I wouldve finetuned more on 2019 but kaggles kernel restrictions hit me pretty hard(no local machine :/).\nMy final ensemble was using the b3(best kappa+loss) and b4/5(all) so 8 models in total.
\n\nNow my thoughts(not technical)\n1. Always read through public kernels/discussions. These are so valuable and provide many insights and can alone probably score a bronze medal. Look through high scoring kernels and see what they do similarly and different and question why. So many people share there training/validation strategies in discussions so it is easy to modify stuff from there.\n2. Try to team up! Not only do teammates raise your score(brought me from bronze -> silver), they can teach so much. I was pretty lucky to team up with @taindow and @rishabhiitbhu (whom i cant thank enough!) because they taught me so much about machine learning, from using k-fold, different training strategies, how to analyze confusion matrices/model statistics, to how to keep track of models and their metrics.
\n\nAnyways, i hope this is helpful to new kagglers and I wish everyone luck in their kaggle adventures!
", "rawMarkdown": "This is just a post to express my thoughts on the competition, and what I learned(this is the first time i dedicated myself to the competition).\n\nI'll just get my approach out of the way:\nMy models alone got a 0.915 private / 0.815 public.\nI originally started this competition by copying @drhabib 's b5 kernel and toying with the thresholds and hyperparams(this got me to 0.798 public lb). Again, I was pretty new to ml/cv so I didn't know how to set up everything.\nI then started on my own approach. Guided by many great kernels from @abhishek and @ratthachat , i slowly began to write my own pipeline and training scripts. \nMy final approach actually became to be training on 2015+2019 with cropping black and using 2019 as validation(probably not the best practice haha). I then trained 3 models each for EfficientNetB2-5 where i checkpointed the model with best loss, best kappa, and the last model. I wouldve finetuned more on 2019 but kaggles kernel restrictions hit me pretty hard(no local machine :/).\nMy final ensemble was using the b3(best kappa+loss) and b4/5(all) so 8 models in total.\n\nNow my thoughts(not technical)\n1. Always read through public kernels/discussions. These are so valuable and provide many insights and can alone probably score a bronze medal. Look through high scoring kernels and see what they do similarly and different and question why. So many people share there training/validation strategies in discussions so it is easy to modify stuff from there.\n2. Try to team up! Not only do teammates raise your score(brought me from bronze -> silver), they can teach so much. I was pretty lucky to team up with @taindow and @rishabhiitbhu (whom i cant thank enough!) because they taught me so much about machine learning, from using k-fold, different training strategies, how to analyze confusion matrices/model statistics, to how to keep track of models and their metrics.\n\nAnyways, i hope this is helpful to new kagglers and I wish everyone luck in their kaggle adventures!", "votes": null }, { "id": "623741", "postDate": "09/11/2019 08:57:42", "content": "On the last point, you really have two great teammates and great friends! ;)
", "rawMarkdown": "On the last point, you really have two great teammates and great friends! ;)", "votes": null }, { "id": "624546", "postDate": "09/12/2019 07:12:26", "content": "Congrats Sidhanth.
", "rawMarkdown": "Congrats Sidhanth.", "votes": null } ], "comments": [ { "id": 623741, "author_name": "ratthachat", "author_url": "", "post_date": "09/11/2019 08:57:42", "content": "On the last point, you really have two great teammates and great friends! ;)
", "votes": null, "replies": [] }, { "id": 624546, "author_name": "manojprabhaakr", "author_url": "", "post_date": "09/12/2019 07:12:26", "content": "Congrats Sidhanth.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "623575": "This is just a post to express my thoughts on the competition, and what I learned(this is the first time i dedicated myself to the competition).\n\nI'll just get my approach out of the way:\nMy models alone got a 0.915 private / 0.815 public.\nI originally started this competition by copying @drhabib 's b5 kernel and toying with the thresholds and hyperparams(this got me to 0.798 public lb). Again, I was pretty new to ml/cv so I didn't know how to set up everything.\nI then started on my own approach. Guided by many great kernels from @abhishek and @ratthachat , i slowly began to write my own pipeline and training scripts. \nMy final approach actually became to be training on 2015+2019 with cropping black and using 2019 as validation(probably not the best practice haha). I then trained 3 models each for EfficientNetB2-5 where i checkpointed the model with best loss, best kappa, and the last model. I wouldve finetuned more on 2019 but kaggles kernel restrictions hit me pretty hard(no local machine :/).\nMy final ensemble was using the b3(best kappa+loss) and b4/5(all) so 8 models in total.\n\nNow my thoughts(not technical)\n1. Always read through public kernels/discussions. These are so valuable and provide many insights and can alone probably score a bronze medal. Look through high scoring kernels and see what they do similarly and different and question why. So many people share there training/validation strategies in discussions so it is easy to modify stuff from there.\n2. Try to team up! Not only do teammates raise your score(brought me from bronze -> silver), they can teach so much. I was pretty lucky to team up with @taindow and @rishabhiitbhu (whom i cant thank enough!) because they taught me so much about machine learning, from using k-fold, different training strategies, how to analyze confusion matrices/model statistics, to how to keep track of models and their metrics.\n\nAnyways, i hope this is helpful to new kagglers and I wish everyone luck in their kaggle adventures!", "623741": "On the last point, you really have two great teammates and great friends! ;)", "624546": "Congrats Sidhanth." }, "source": "meta" }