![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F3492127%2Ff14ba2a956412e823537089b0d2c05d2%2F41.png?generation=1595355611359172&alt=media\")
Thank you to the organizers for this interesting competition!\nCongrats to all the winners and participants.
\n\nIn fact, I haven't tried a lot.\nWe realized the way too late.
\n\nHowever it can be helpful for someone, so simply write it down.\n(I did other experiments - pretrained other models, optimizers, etc. - but I'm going to write down the main issues.)
\n\nIt was good to do simple experiments, but it was hard to exceed 0.930 lb.
\n\nAnd then I trained b7.\nIt had the best performance on the about epoch 30 and achieved 0.933 lb.\n(I haven't tested many epochs due to limited GPU resources.)
\n\nLearned just before the close of the competition and achieved 0.930 lb at about 28 epoch.
\n\nflip augmentation\nrotate +- 10 (just B6)\ncutmix (our teammate - B3, @kani23 )
\n\nWhen I applied rotate and cutmix, the validation score was a little higher.\nlike this:\n[RESULT]: Train. Epoch: 21, summary_loss: 0.82028, final_score: 0.87635, time: 4017.75592\n[RESULT]: Val. Epoch: 21, summary_loss: 0.75268, final_score: 0.90341, time: 325.94613
Above 0.930 lb, it didn't work in public lb.\nbut in private, Actually, there was a very big difference.\nEfficientNet-B7 + no TTA : 0.932 lb / (private 0.916)\nEfficientNet-B7 + TTA : 0.932 lb / (private 0.920)
Unfortunately I ensemble with the results of not applying TTA.\nSo, it didn't score better than my single model.\nš
\n\n
Anyway, I learned a lot and I will try harder in the future.\nThanks!
\n\np.s. Thanks to our teammate @kani23,\np.s.2. @loveall - You can not use our results for your paper job.\nbecause you have contributed nothing and please don't take someone else's kernel anymore.
", "messages": [ { "id": "937694", "postDate": "07/21/2020 06:06:10", "content": "Thank you to the organizers for this interesting competition!\nCongrats to all the winners and participants.
\n\nIn fact, I haven't tried a lot.\nWe realized the way too late.
\n\nHowever it can be helpful for someone, so simply write it down.\n(I did other experiments - pretrained other models, optimizers, etc. - but I'm going to write down the main issues.)
\n\nIt was good to do simple experiments, but it was hard to exceed 0.930 lb.
\n\nAnd then I trained b7.\nIt had the best performance on the about epoch 30 and achieved 0.933 lb.\n(I haven't tested many epochs due to limited GPU resources.)
\n\nLearned just before the close of the competition and achieved 0.930 lb at about 28 epoch.
\n\nflip augmentation\nrotate +- 10 (just B6)\ncutmix (our teammate - B3, @kani23 )
\n\nWhen I applied rotate and cutmix, the validation score was a little higher.\nlike this:\n[RESULT]: Train. Epoch: 21, summary_loss: 0.82028, final_score: 0.87635, time: 4017.75592\n[RESULT]: Val. Epoch: 21, summary_loss: 0.75268, final_score: 0.90341, time: 325.94613
Above 0.930 lb, it didn't work in public lb.\nbut in private, Actually, there was a very big difference.\nEfficientNet-B7 + no TTA : 0.932 lb / (private 0.916)\nEfficientNet-B7 + TTA : 0.932 lb / (private 0.920)
Unfortunately I ensemble with the results of not applying TTA.\nSo, it didn't score better than my single model.\nš
\n\n
Anyway, I learned a lot and I will try harder in the future.\nThanks!
\n\np.s. Thanks to our teammate @kani23,\np.s.2. @loveall - You can not use our results for your paper job.\nbecause you have contributed nothing and please don't take someone else's kernel anymore.
", "rawMarkdown": "Thank you to the organizers for this interesting competition!\nCongrats to all the winners and participants.\n\nIn fact, I haven't tried a lot.\nWe realized the way too late.\n\nHowever it can be helpful for someone, so simply write it down.\n(I did other experiments - pretrained other models, optimizers, etc. - but I'm going to write down the main issues.)\n\n# EfficientNet-B0, EfficientNet-B1, EfficientNet-B2\nIt was good to do simple experiments, but it was hard to exceed 0.930 lb.\n\n# EfficientNet-B7\nAnd then I trained b7.\nIt had the best performance on the about epoch 30 and achieved 0.933 lb.\n(I haven't tested many epochs due to limited GPU resources.)\n\n# EfficientNet-B6\nLearned just before the close of the competition and achieved 0.930 lb at about 28 epoch.\n\n# Agumentation\nflip augmentation\nrotate +- 10 (just B6)\ncutmix (our teammate - B3, @[kani23](https://www.kaggle.com/kani23) )\n\nWhen I applied rotate and cutmix, the validation score was a little higher.\nlike this:\n`[RESULT]: Train. Epoch: 21, summary_loss: 0.82028, final_score: 0.87635, time: 4017.75592\n[RESULT]: Val. Epoch: 21, summary_loss: 0.75268, final_score: 0.90341, time: 325.94613`\n\n# TTA\nAbove 0.930 lb, it didn't work in public lb.\nbut in private, Actually, there was a very big difference.\n`EfficientNet-B7 + no TTA : 0.932 lb / (private 0.916)`\n`EfficientNet-B7 + TTA : 0.932 lb / (private 0.920)`\n\nUnfortunately I ensemble with the results of not applying TTA.\nSo, it didn't score better than my single model.\nš \n\n\n\n \nAnyway, I learned a lot and I will try harder in the future.\nThanks!\n\np.s. Thanks to our teammate @[kani23](https://www.kaggle.com/kani23),\np.s.2. [@loveall ](https://www.kaggle.com/loveall) - You can not use our results for your paper job.\nbecause you have contributed nothing and please don't take someone else's kernel anymore.", "votes": null }, { "id": "938409", "postDate": "07/21/2020 14:05:03", "content": "@piantic - what type of hardware did you use for this competition? You said you had limited GPU resources so just curious to know.
", "rawMarkdown": "piantic - what type of hardware did you use for this competition? You said you had limited GPU resources so just curious to know.", "votes": null }, { "id": "938437", "postDate": "07/21/2020 14:27:48", "content": "Initially I used TPU However, the Kaggle TPU was limited. \nLater, I used my local 2080Ti.\nAt the end of the competition, V100 was used for a few days . š
", "rawMarkdown": "Initially I used TPU However, the Kaggle TPU was limited. \nLater, I used my local 2080Ti.\nAt the end of the competition, V100 was used for a few days . š", "votes": null } ], "comments": [ { "id": 938409, "author_name": "rdizzl3", "author_url": "", "post_date": "07/21/2020 14:05:03", "content": "@piantic - what type of hardware did you use for this competition? You said you had limited GPU resources so just curious to know.
", "votes": null, "replies": [ { "id": 938437, "author_name": "piantic", "author_url": "", "post_date": "07/21/2020 14:27:48", "content": "Initially I used TPU However, the Kaggle TPU was limited. \nLater, I used my local 2080Ti.\nAt the end of the competition, V100 was used for a few days . š
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "937694": "Thank you to the organizers for this interesting competition!\nCongrats to all the winners and participants.\n\nIn fact, I haven't tried a lot.\nWe realized the way too late.\n\nHowever it can be helpful for someone, so simply write it down.\n(I did other experiments - pretrained other models, optimizers, etc. - but I'm going to write down the main issues.)\n\n# EfficientNet-B0, EfficientNet-B1, EfficientNet-B2\nIt was good to do simple experiments, but it was hard to exceed 0.930 lb.\n\n# EfficientNet-B7\nAnd then I trained b7.\nIt had the best performance on the about epoch 30 and achieved 0.933 lb.\n(I haven't tested many epochs due to limited GPU resources.)\n\n# EfficientNet-B6\nLearned just before the close of the competition and achieved 0.930 lb at about 28 epoch.\n\n# Agumentation\nflip augmentation\nrotate +- 10 (just B6)\ncutmix (our teammate - B3, @[kani23](https://www.kaggle.com/kani23) )\n\nWhen I applied rotate and cutmix, the validation score was a little higher.\nlike this:\n`[RESULT]: Train. Epoch: 21, summary_loss: 0.82028, final_score: 0.87635, time: 4017.75592\n[RESULT]: Val. Epoch: 21, summary_loss: 0.75268, final_score: 0.90341, time: 325.94613`\n\n# TTA\nAbove 0.930 lb, it didn't work in public lb.\nbut in private, Actually, there was a very big difference.\n`EfficientNet-B7 + no TTA : 0.932 lb / (private 0.916)`\n`EfficientNet-B7 + TTA : 0.932 lb / (private 0.920)`\n\nUnfortunately I ensemble with the results of not applying TTA.\nSo, it didn't score better than my single model.\nš \n\n\n\n \nAnyway, I learned a lot and I will try harder in the future.\nThanks!\n\np.s. Thanks to our teammate @[kani23](https://www.kaggle.com/kani23),\np.s.2. [@loveall ](https://www.kaggle.com/loveall) - You can not use our results for your paper job.\nbecause you have contributed nothing and please don't take someone else's kernel anymore.", "938409": "piantic - what type of hardware did you use for this competition? You said you had limited GPU resources so just curious to know.", "938437": "Initially I used TPU However, the Kaggle TPU was limited. \nLater, I used my local 2080Ti.\nAt the end of the competition, V100 was used for a few days . š" }, "source": "meta" }