It was my first time to join kaggle competition and luckily I got a silver medal. Thank you for everyone and the host of competition. Specially, I want to say thank you for khyeh0719 who provided this notebook as a good baseline for me. I learnt how to build a pipeline by PyTorch.
\nOne thing from my solution that I want to recap is the SCE (Symmetric Cross Entropy) loss function I used in my model. It actually worked for me. It pushed my baseline from 0.8950 to 0.9014 which is simple but efficient way, you only need to switch the loss function. The paper and code you can find in the links.
It seems no one mention it yet so I write this discussion to share SCE loss function. If it helps you, please give me a vote😄😄. I hope it will not waste your time. If you have any further question, feel free to comment below and I love to discuss with you.
", "messages": [ { "id": "1211186", "postDate": "02/20/2021 03:27:42", "content": "It was my first time to join kaggle competition and luckily I got a silver medal. Thank you for everyone and the host of competition. Specially, I want to say thank you for khyeh0719 who provided this notebook as a good baseline for me. I learnt how to build a pipeline by PyTorch.
\nOne thing from my solution that I want to recap is the SCE (Symmetric Cross Entropy) loss function I used in my model. It actually worked for me. It pushed my baseline from 0.8950 to 0.9014 which is simple but efficient way, you only need to switch the loss function. The paper and code you can find in the links.
It seems no one mention it yet so I write this discussion to share SCE loss function. If it helps you, please give me a vote😄😄. I hope it will not waste your time. If you have any further question, feel free to comment below and I love to discuss with you.
", "rawMarkdown": "It was my first time to join kaggle competition and luckily I got a silver medal. Thank you for everyone and the host of competition. Specially, I want to say thank you for [khyeh0719](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug) who provided this notebook as a good baseline for me. I learnt how to build a pipeline by PyTorch.\n\nOne thing from my solution that I want to recap is the SCE (Symmetric Cross Entropy) loss function I used in my model. It actually worked for me. It pushed my baseline from `0.8950` to `0.9014` which is simple but efficient way, you only need to switch the loss function. The [paper](https://openaccess.thecvf.com/content_ICCV_2019/html/Wang_Symmetric_Cross_Entropy_for_Robust_Learning_With_Noisy_Labels_ICCV_2019_paper.html) and [code](https://github.com/HanxunH/SCELoss-Reproduce) you can find in the links.\n\nIt seems no one mention it yet so I write this discussion to share SCE loss function. If it helps you, please give me a vote😄😄. I hope it will not waste your time. If you have any further question, feel free to comment below and I love to discuss with you.", "votes": null }, { "id": "1211848", "postDate": "02/20/2021 16:02:47", "content": "I have never seen this symmetrized cross entropy in machine learning loss context before, but I really like the idea and I am happy that it gave you such a boost… Thanks for sharing this information and references.
", "rawMarkdown": "I have never seen this symmetrized cross entropy in machine learning loss context before, but I really like the idea and I am happy that it gave you such a boost... Thanks for sharing this information and references.", "votes": null }, { "id": "1212611", "postDate": "02/21/2021 11:57:44", "content": "Yeah, the idea is good and it is easy to implement. Hope it helps you too.
", "rawMarkdown": "Yeah, the idea is good and it is easy to implement. Hope it helps you too.", "votes": null } ], "comments": [ { "id": 1211848, "author_name": "morodertobias", "author_url": "", "post_date": "02/20/2021 16:02:47", "content": "I have never seen this symmetrized cross entropy in machine learning loss context before, but I really like the idea and I am happy that it gave you such a boost… Thanks for sharing this information and references.
", "votes": null, "replies": [ { "id": 1212611, "author_name": "dongjai04", "author_url": "", "post_date": "02/21/2021 11:57:44", "content": "Yeah, the idea is good and it is easy to implement. Hope it helps you too.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1211186": "It was my first time to join kaggle competition and luckily I got a silver medal. Thank you for everyone and the host of competition. Specially, I want to say thank you for [khyeh0719](https://www.kaggle.com/khyeh0719/pytorch-efficientnet-baseline-train-amp-aug) who provided this notebook as a good baseline for me. I learnt how to build a pipeline by PyTorch.\n\nOne thing from my solution that I want to recap is the SCE (Symmetric Cross Entropy) loss function I used in my model. It actually worked for me. It pushed my baseline from `0.8950` to `0.9014` which is simple but efficient way, you only need to switch the loss function. The [paper](https://openaccess.thecvf.com/content_ICCV_2019/html/Wang_Symmetric_Cross_Entropy_for_Robust_Learning_With_Noisy_Labels_ICCV_2019_paper.html) and [code](https://github.com/HanxunH/SCELoss-Reproduce) you can find in the links.\n\nIt seems no one mention it yet so I write this discussion to share SCE loss function. If it helps you, please give me a vote😄😄. I hope it will not waste your time. If you have any further question, feel free to comment below and I love to discuss with you.", "1211848": "I have never seen this symmetrized cross entropy in machine learning loss context before, but I really like the idea and I am happy that it gave you such a boost... Thanks for sharing this information and references.", "1212611": "Yeah, the idea is good and it is easy to implement. Hope it helps you too." }, "source": "meta" }