This is the first competition i participate about computer vision.As a rookie, winning a silver medal is definitely a surprise for me. I have to say, luck is important more or less since the data is dirty and the score of many people is close.
\nThanks
\nNotebooks from Gabriel Prado help me a lot.It saves time for me.Actually,lb score 0.901 is not low,so i got that some methods in his work is effective.
\nhttps://www.kaggle.com/capiru/cassavanet-starter-easy-gpu-tpu-cv-0-9
\nhttps://www.kaggle.com/capiru/cassavanet-inference-tta-easy-submission
\nAnd thanks for my teammates's excellent work.
Useful attempt
\nDepening on his work,i continued trying some methods.Some of them are always useful,but others is not so sure.
\nensemble for few models: I tried efficinets resnext50 vit. The b4,the best,scores 0.903 in lb,more than other.Ensemble these brings a better score.
\nuse extra dataset: Using 2019 dataset and making soft labels for extra and test images may not bring improvement but at least not hurt,and it makes the model more robust.
Questions
\nIs Knowledge Distillation helpful? I notice that the champion in 2019 used it.I tried,and my cv improves a little for some folds.But i doubt whether this little improvement will help and i was affraid that the other models was just 'repeating' the teacher model(i used a b4 with best cv for teacher),so i even didn't upload it to lb.
\nWhen can TPU run faster than GPU? I finds that they run nearly as fast as each other on the same batch_size.
\nWhy RamdomCrop or RamdomResizeCrop didn't hurt the prediction in test? In my opinion, it's useful in trainging, but test images been cropped ramdomly really wolud't hurt?And i notice that the acquiescent crop size is 0.08-1,which means perhaps most of the image is lost.
CLASStorchvision.transforms.RandomResizedCrop(size, scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), interpolation=2)
This is the first competition i participate about computer vision.As a rookie, winning a silver medal is definitely a surprise for me. I have to say, luck is important more or less since the data is dirty and the score of many people is close.
\nThanks
\nNotebooks from Gabriel Prado help me a lot.It saves time for me.Actually,lb score 0.901 is not low,so i got that some methods in his work is effective.
\nhttps://www.kaggle.com/capiru/cassavanet-starter-easy-gpu-tpu-cv-0-9
\nhttps://www.kaggle.com/capiru/cassavanet-inference-tta-easy-submission
\nAnd thanks for my teammates's excellent work.
Useful attempt
\nDepening on his work,i continued trying some methods.Some of them are always useful,but others is not so sure.
\nensemble for few models: I tried efficinets resnext50 vit. The b4,the best,scores 0.903 in lb,more than other.Ensemble these brings a better score.
\nuse extra dataset: Using 2019 dataset and making soft labels for extra and test images may not bring improvement but at least not hurt,and it makes the model more robust.
Questions
\nIs Knowledge Distillation helpful? I notice that the champion in 2019 used it.I tried,and my cv improves a little for some folds.But i doubt whether this little improvement will help and i was affraid that the other models was just 'repeating' the teacher model(i used a b4 with best cv for teacher),so i even didn't upload it to lb.
\nWhen can TPU run faster than GPU? I finds that they run nearly as fast as each other on the same batch_size.
\nWhy RamdomCrop or RamdomResizeCrop didn't hurt the prediction in test? In my opinion, it's useful in trainging, but test images been cropped ramdomly really wolud't hurt?And i notice that the acquiescent crop size is 0.08-1,which means perhaps most of the image is lost.
CLASStorchvision.transforms.RandomResizedCrop(size, scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), interpolation=2)
Congrats on your a first silver medal. :) @jsrdcht
", "rawMarkdown": "Congrats on your a first silver medal. :) @jsrdcht", "votes": null }, { "id": "1210282", "postDate": "02/19/2021 10:14:05", "content": "Thanks! I have to say, this competition is favorable for people who does't have much experience.
", "rawMarkdown": "Thanks! I have to say, this competition is favorable for people who does't have much experience.", "votes": null } ], "comments": [ { "id": 1210272, "author_name": "piantic", "author_url": "", "post_date": "02/19/2021 10:03:35", "content": "Congrats on your a first silver medal. :) @jsrdcht
", "votes": null, "replies": [ { "id": 1210282, "author_name": "jsrdcht", "author_url": "", "post_date": "02/19/2021 10:14:05", "content": "Thanks! I have to say, this competition is favorable for people who does't have much experience.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1210268": "This is the first competition i participate about computer vision.As a rookie, winning a silver medal is definitely a surprise for me. I have to say, luck is important more or less since the data is dirty and the score of many people is close.\n\n**Thanks**\nNotebooks from *Gabriel Prado* help me a lot.It saves time for me.Actually,lb score 0.901 is not low,so i got that some methods in his work is effective.\nhttps://www.kaggle.com/capiru/cassavanet-starter-easy-gpu-tpu-cv-0-9\nhttps://www.kaggle.com/capiru/cassavanet-inference-tta-easy-submission\nAnd thanks for my teammates's excellent work.\n\n**Useful attempt**\nDepening on his work,i continued trying some methods.Some of them are always useful,but others is not so sure.\nensemble for few models: I tried efficinets resnext50 vit. The b4,the best,scores 0.903 in lb,more than other.Ensemble these brings a better score.\nuse extra dataset: Using 2019 dataset and making soft labels for extra and test images may not bring improvement but at least not hurt,and it makes the model more robust.\n\n**Questions**\nIs Knowledge Distillation helpful? I notice that the champion in 2019 used it.I tried,and my cv improves a little for some folds.But i doubt whether this little improvement will help and i was affraid that the other models was just 'repeating' the teacher model(i used a b4 with best cv for teacher),so i even didn't upload it to lb.\nWhen can TPU run faster than GPU? I finds that they run nearly as fast as each other on the same batch_size.\nWhy RamdomCrop or RamdomResizeCrop didn't hurt the prediction in test? In my opinion, it's useful in trainging, but test images been cropped ramdomly really wolud't hurt?And i notice that the acquiescent crop size is 0.08-1,which means perhaps most of the image is lost.\n\n`CLASStorchvision.transforms.RandomResizedCrop(size, scale=(0.08, 1.0), ratio=(0.75, 1.3333333333333333), interpolation=2)`", "1210272": "Congrats on your a first silver medal. :) @jsrdcht", "1210282": "Thanks! I have to say, this competition is favorable for people who does't have much experience." }, "source": "meta" }