{ "id": 242050, "title": "We Misused ImageNet-21k pretrained Swin-transformer", "url": "/competitions/semi-inat-2021/discussion/242050", "author_name": "", "post_date": "2021-05-27T08:58:49.922475800Z", "votes": 1, "comment_count": 2, "views": 0, "content": "

Dear hosts and all competitors:

\n

I am with the team simple_test_test. In our solution, we have used a swin-transformer as our backbone. The library timm [https://github.com/rwightman/pytorch-image-models] marked the pretrained swin-transformer model as pretrained on ImageNet 1k, but it is pretrained on ImageNet 21k and then finetuned on ImageNet 1k [https://github.com/rwightman/pytorch-image-models/blob/9c78de8c024bff0acc68b044dfb935366c6185dc/timm/models/swin_transformer.py#L49]. So our top submissions have wrongly involved ImageNet 21k pretrained backbone by this innocent mistake. We will select submissions that only used ImageNet 1k pretrained backbone as our final submission, instead of top submissions on the public leader board. We hope this will not be against the rules.

\n

Sincerely apologize to hosts and other competitors

", "messages": [ { "id": "1324834", "postDate": "05/27/2021 08:58:49", "content": "

Dear hosts and all competitors:

\n

I am with the team simple_test_test. In our solution, we have used a swin-transformer as our backbone. The library timm [https://github.com/rwightman/pytorch-image-models] marked the pretrained swin-transformer model as pretrained on ImageNet 1k, but it is pretrained on ImageNet 21k and then finetuned on ImageNet 1k [https://github.com/rwightman/pytorch-image-models/blob/9c78de8c024bff0acc68b044dfb935366c6185dc/timm/models/swin_transformer.py#L49]. So our top submissions have wrongly involved ImageNet 21k pretrained backbone by this innocent mistake. We will select submissions that only used ImageNet 1k pretrained backbone as our final submission, instead of top submissions on the public leader board. We hope this will not be against the rules.

\n

Sincerely apologize to hosts and other competitors

", "rawMarkdown": "Dear hosts and all competitors:\n\nI am with the team simple_test_test. In our solution, we have used a swin-transformer as our backbone. The library timm [https://github.com/rwightman/pytorch-image-models] marked the pretrained swin-transformer model as pretrained on ImageNet 1k, but it is pretrained on ImageNet 21k and then finetuned on ImageNet 1k [https://github.com/rwightman/pytorch-image-models/blob/9c78de8c024bff0acc68b044dfb935366c6185dc/timm/models/swin_transformer.py#L49]. So our top submissions have wrongly involved ImageNet 21k pretrained backbone by this innocent mistake. We will select submissions that only used ImageNet 1k pretrained backbone as our final submission, instead of top submissions on the public leader board. We hope this will not be against the rules. \n\nSincerely apologize to hosts and other competitors", "votes": null }, { "id": "1325279", "postDate": "05/27/2021 16:11:46", "content": "

Thanks for letting us know. I can delete those incorrect submissions for you. Could you send me an email (jcsu@cs.umass.edu) saying which submission you want to discard/keep? Thank you!

", "rawMarkdown": "Thanks for letting us know. I can delete those incorrect submissions for you. Could you send me an email (jcsu@cs.umass.edu) saying which submission you want to discard/keep? Thank you!", "votes": null }, { "id": "1331672", "postDate": "06/01/2021 16:11:54", "content": "

Hi @simple_test_test, I did not receive your email. Since I don't have your email, I cannot contact you. Could you email me or leave your contact here? Thank you!

", "rawMarkdown": "Hi @simple_test_test, I did not receive your email. Since I don't have your email, I cannot contact you. Could you email me or leave your contact here? Thank you!", "votes": null } ], "comments": [ { "id": 1325279, "author_name": "jcfredsu", "author_url": "", "post_date": "05/27/2021 16:11:46", "content": "

Thanks for letting us know. I can delete those incorrect submissions for you. Could you send me an email (jcsu@cs.umass.edu) saying which submission you want to discard/keep? Thank you!

", "votes": null, "replies": [] }, { "id": 1331672, "author_name": "jcfredsu", "author_url": "", "post_date": "06/01/2021 16:11:54", "content": "

Hi @simple_test_test, I did not receive your email. Since I don't have your email, I cannot contact you. Could you email me or leave your contact here? Thank you!

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1324834": "Dear hosts and all competitors:\n\nI am with the team simple_test_test. In our solution, we have used a swin-transformer as our backbone. The library timm [https://github.com/rwightman/pytorch-image-models] marked the pretrained swin-transformer model as pretrained on ImageNet 1k, but it is pretrained on ImageNet 21k and then finetuned on ImageNet 1k [https://github.com/rwightman/pytorch-image-models/blob/9c78de8c024bff0acc68b044dfb935366c6185dc/timm/models/swin_transformer.py#L49]. So our top submissions have wrongly involved ImageNet 21k pretrained backbone by this innocent mistake. We will select submissions that only used ImageNet 1k pretrained backbone as our final submission, instead of top submissions on the public leader board. We hope this will not be against the rules. \n\nSincerely apologize to hosts and other competitors", "1325279": "Thanks for letting us know. I can delete those incorrect submissions for you. Could you send me an email (jcsu@cs.umass.edu) saying which submission you want to discard/keep? Thank you!", "1331672": "Hi @simple_test_test, I did not receive your email. Since I don't have your email, I cannot contact you. Could you email me or leave your contact here? Thank you!" }, "source": "meta" }