First of all thanks to the Kaggle community for active participation in the discussions and notebooks. A special mention to the pipeline set by @cdeotte in his great notebook Triple Stratified KFold with TFRecords (https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). It helped us (https://www.kaggle.com/vikrant06 and me) quickly try out multiple configurations of EffNet. What also seemed to work was the 'noisy-student' Effnet weights.
\nWe used 3 folds including the data of 2018 and 2019 on 384x384 images along with brightness and contrast augmentation on B5, B6 and B7 (imagenet and noisy student weights for each) ensembled using power averaging with the power of 2.
\nThe resultant ensemble when ensembled (weighted average) against the top 4 public leaderboard solutions (which were ensembled using power averaging with the power of 2) gave us Public LB score of 0.9644 and private LB score of 0.9458.
", "messages": [ { "id": 977223, "postDate": "2020-08-19T11:02:36.773Z", "content": "First of all thanks to the Kaggle community for active participation in the discussions and notebooks. A special mention to the pipeline set by @cdeotte in his great notebook Triple Stratified KFold with TFRecords (https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). It helped us (https://www.kaggle.com/vikrant06 and me) quickly try out multiple configurations of EffNet. What also seemed to work was the 'noisy-student' Effnet weights.
\nWe used 3 folds including the data of 2018 and 2019 on 384x384 images along with brightness and contrast augmentation on B5, B6 and B7 (imagenet and noisy student weights for each) ensembled using power averaging with the power of 2.
\nThe resultant ensemble when ensembled (weighted average) against the top 4 public leaderboard solutions (which were ensembled using power averaging with the power of 2) gave us Public LB score of 0.9644 and private LB score of 0.9458.
", "rawMarkdown": "First of all thanks to the Kaggle community for active participation in the discussions and notebooks. A special mention to the pipeline set by @cdeotte in his great notebook Triple Stratified KFold with TFRecords (https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). It helped us (https://www.kaggle.com/vikrant06 and me) quickly try out multiple configurations of EffNet. What also seemed to work was the 'noisy-student' Effnet weights. \n\nWe used 3 folds including the data of 2018 and 2019 on 384x384 images along with brightness and contrast augmentation on B5, B6 and B7 (imagenet and noisy student weights for each) ensembled using power averaging with the power of 2. \n\nThe resultant ensemble when ensembled (weighted average) against the top 4 public leaderboard solutions (which were ensembled using power averaging with the power of 2) gave us Public LB score of 0.9644 and private LB score of 0.9458.", "votes": 10 }, { "id": 986391, "postDate": "2020-08-26T12:53:58.427Z", "content": "Can you please upload your code on Github or share your notebook kernels.
\nThanks
great
", "rawMarkdown": "great" }, { "id": 981058, "postDate": "2020-08-22T06:19:32.850Z", "content": "Congrats!
", "rawMarkdown": "Congrats!" }, { "id": 978265, "postDate": "2020-08-20T04:12:47.883Z", "content": "Congratulations!
", "rawMarkdown": "Congratulations!", "replies": [ { "id": 978554, "postDate": "2020-08-20T08:49:08.350Z", "content": "Thank you so much
", "rawMarkdown": "Thank you so much" } ] }, { "id": 977356, "postDate": "2020-08-19T12:39:08.767Z", "content": "Congrats . Did you run B7 with 3 fold in one session as I was not able to . if yes how you did it.
", "rawMarkdown": "Congrats . Did you run B7 with 3 fold in one session as I was not able to . if yes how you did it.", "replies": [ { "id": 977467, "postDate": "2020-08-19T13:50:34.493Z", "content": "Me neither. I had memory issues every time I tried.
", "rawMarkdown": "Me neither. I had memory issues every time I tried." }, { "id": 977500, "postDate": "2020-08-19T14:17:49.687Z", "content": "Yes, I was able to run 16 epochs on B7. Batch size 32 using @cdeotte 's notebook in 3 folds for 384x384 images of 2018, 2019 and 2020. Otherwise try with 256x256 once.
\nYou will have to add efn.EfficientNetB7 in EFNS list.
", "rawMarkdown": "Yes, I was able to run 16 epochs on B7. Batch size 32 using @cdeotte 's notebook in 3 folds for 384x384 images of 2018, 2019 and 2020. Otherwise try with 256x256 once.\n\nYou will have to add efn.EfficientNetB7 in EFNS list." } ] }, { "id": 979797, "postDate": "2020-08-21T05:57:44.790Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 986391, "author_name": "Appy Patel", "author_url": "", "post_date": "2020-08-26T12:53:58.427000", "content": "Can you please upload your code on Github or share your notebook kernels.
\nThanks
great
", "votes": 0, "replies": [] }, { "id": 981058, "author_name": "Abishek Sudarshan", "author_url": "", "post_date": "2020-08-22T06:19:32.850000", "content": "Congrats!
", "votes": 0, "replies": [] }, { "id": 978265, "author_name": "Alin Cijov", "author_url": "", "post_date": "2020-08-20T04:12:47.883000", "content": "Congratulations!
", "votes": 0, "replies": [ { "id": 978554, "author_name": "Tejas Kotha", "author_url": "", "post_date": "2020-08-20T08:49:08.350000", "content": "Thank you so much
", "votes": 0, "replies": [] } ] }, { "id": 977356, "author_name": "Rajnish Chauhan", "author_url": "", "post_date": "2020-08-19T12:39:08.767000", "content": "Congrats . Did you run B7 with 3 fold in one session as I was not able to . if yes how you did it.
", "votes": 0, "replies": [ { "id": 977467, "author_name": "Santiago Viquez", "author_url": "", "post_date": "2020-08-19T13:50:34.493000", "content": "Me neither. I had memory issues every time I tried.
", "votes": 0, "replies": [] }, { "id": 977500, "author_name": "Tejas Kotha", "author_url": "", "post_date": "2020-08-19T14:17:49.687000", "content": "Yes, I was able to run 16 epochs on B7. Batch size 32 using @cdeotte 's notebook in 3 folds for 384x384 images of 2018, 2019 and 2020. Otherwise try with 256x256 once.
\nYou will have to add efn.EfficientNetB7 in EFNS list.
", "votes": 0, "replies": [] } ] }, { "id": 979797, "author_name": "", "author_url": "", "post_date": "2020-08-21T05:57:44.790000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "977223": "First of all thanks to the Kaggle community for active participation in the discussions and notebooks. A special mention to the pipeline set by @cdeotte in his great notebook Triple Stratified KFold with TFRecords (https://www.kaggle.com/cdeotte/triple-stratified-kfold-with-tfrecords). It helped us (https://www.kaggle.com/vikrant06 and me) quickly try out multiple configurations of EffNet. What also seemed to work was the 'noisy-student' Effnet weights. \n\nWe used 3 folds including the data of 2018 and 2019 on 384x384 images along with brightness and contrast augmentation on B5, B6 and B7 (imagenet and noisy student weights for each) ensembled using power averaging with the power of 2. \n\nThe resultant ensemble when ensembled (weighted average) against the top 4 public leaderboard solutions (which were ensembled using power averaging with the power of 2) gave us Public LB score of 0.9644 and private LB score of 0.9458.", "986391": "Can you please upload your code on Github or share your notebook kernels.\nThanks", "981338": "great", "981058": "Congrats!", "978265": "Congratulations!", "977356": "Congrats . Did you run B7 with 3 fold in one session as I was not able to . if yes how you did it.", "979797": "" } }