This is my first competition and I am so luckily get silver medal for the first time. I just entered this competition 2 weeks ago (too late) but I learned a lot of things about code, train model, tricks from discussion topics and public notebooks . I got the silver medal by using extra images from 2019 competition, train single model with (5 fold CV strategy, Cutmix + light augmentation, bi-tempered loss with label smoothing 0.2, cosine annealing scheduler with warm up, Adam optimizer, AMP), ensemble model (EfficientNet b4 ns, EfficientNet b5 ns, SEResNeXt50, SEResNeXt101, ViT patch 16 base 384, Resnet 200d, ResNeXt50). After using ensemble model to create pseudo labels in extra image in 2019 competition and soft labels for knowledge distillation (0.7gt + 0.3soft_label) source in 2019 + 2020 train and train 2 new model (EfficientNet b5 ns, SEResNeXt50) on both new labels for extra and train images. Final ensemble all models with forward selection (I tried combine it with beam search but both same OOF CV) and inference with no TTA.
\nThis is a good competition to learn a lot of things to handle noisy data since a lot of the real world data can be mislabeled. And also, can someone share with me about previous good competitions to learn?. Thank you very very very much!
\nThank everyone and see you again in other competitions!
This is my first competition and I am so luckily get silver medal for the first time. I just entered this competition 2 weeks ago (too late) but I learned a lot of things about code, train model, tricks from discussion topics and public notebooks . I got the silver medal by using extra images from 2019 competition, train single model with (5 fold CV strategy, Cutmix + light augmentation, bi-tempered loss with label smoothing 0.2, cosine annealing scheduler with warm up, Adam optimizer, AMP), ensemble model (EfficientNet b4 ns, EfficientNet b5 ns, SEResNeXt50, SEResNeXt101, ViT patch 16 base 384, Resnet 200d, ResNeXt50). After using ensemble model to create pseudo labels in extra image in 2019 competition and soft labels for knowledge distillation (0.7gt + 0.3soft_label) source in 2019 + 2020 train and train 2 new model (EfficientNet b5 ns, SEResNeXt50) on both new labels for extra and train images. Final ensemble all models with forward selection (I tried combine it with beam search but both same OOF CV) and inference with no TTA.
\nThis is a good competition to learn a lot of things to handle noisy data since a lot of the real world data can be mislabeled. And also, can someone share with me about previous good competitions to learn?. Thank you very very very much!
\nThank everyone and see you again in other competitions!
Nice job!. Congrats on a first your silver medal. @hungkhoi
", "rawMarkdown": "Nice job!. Congrats on a first your silver medal. @hungkhoi", "votes": null }, { "id": "1210291", "postDate": "02/19/2021 10:17:52", "content": "Thank you for sharing your notebooks!
", "rawMarkdown": "Thank you for sharing your notebooks!", "votes": null } ], "comments": [ { "id": 1210279, "author_name": "piantic", "author_url": "", "post_date": "02/19/2021 10:09:25", "content": "Nice job!. Congrats on a first your silver medal. @hungkhoi
", "votes": null, "replies": [ { "id": 1210291, "author_name": "hungkhoi", "author_url": "", "post_date": "02/19/2021 10:17:52", "content": "Thank you for sharing your notebooks!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1209816": "This is my first competition and I am so luckily get silver medal for the first time. I just entered this competition 2 weeks ago (too late) but I learned a lot of things about code, train model, tricks from discussion topics and public notebooks . I got the silver medal by using extra images from 2019 competition, train single model with (5 fold CV strategy, Cutmix + light augmentation, bi-tempered loss with label smoothing 0.2, cosine annealing scheduler with warm up, Adam optimizer, AMP), ensemble model (EfficientNet b4 ns, EfficientNet b5 ns, SEResNeXt50, SEResNeXt101, ViT patch 16 base 384, Resnet 200d, ResNeXt50). After using ensemble model to create pseudo labels in extra image in 2019 competition and soft labels for knowledge distillation (0.7*gt + 0.3*soft_label) [source](https://www.kaggle.com/c/plant-pathology-2020-fgvc7/discussion/154056) in 2019 + 2020 train and train 2 new model (EfficientNet b5 ns, SEResNeXt50) on both new labels for extra and train images. Final ensemble all models with forward selection (I tried combine it with beam search but both same OOF CV) and inference with no TTA.\n\n This is a good competition to learn a lot of things to handle noisy data since a lot of the real world data can be mislabeled. And also, can someone share with me about previous good competitions to learn?. Thank you very very very much! \n\n Thank everyone and see you again in other competitions!", "1210279": "Nice job!. Congrats on a first your silver medal. @hungkhoi", "1210291": "Thank you for sharing your notebooks!" }, "source": "meta" }