{ "id": 110689, "title": "Publicly, freely available pretrained models", "url": "/competitions/rsna-intracranial-hemorrhage-detection/discussion/110689", "author_name": "Mindy X", "post_date": "2019-09-30T11:07:29.882000", "votes": 11, "comment_count": 9, "views": 0, "content": "

Is it allowed to use publicly, freely available pretrained models trained on non-public dataset? Thanks!

", "messages": [ { "id": 636920, "postDate": "2019-09-30T11:07:29.883Z", "content": "

Is it allowed to use publicly, freely available pretrained models trained on non-public dataset? Thanks!

", "rawMarkdown": "Is it allowed to use publicly, freely available pretrained models trained on non-public dataset? Thanks!", "votes": 11 }, { "id": 640831, "postDate": "2019-10-04T09:01:14.157Z", "content": "

@juliaelliott Can you confirm if we are able to use https://github.com/facebookresearch/WSL-Images?

\n\n

This model is pre-trained in a weakly supervised fashion on a dataset that is not available to the public (a bunch of instagram images), but is then finetuned on the public ImageNet1k like all the other common pre-trained models.

\n\n

Would this violate the rules?

", "rawMarkdown": "@juliaelliott Can you confirm if we are able to use https://github.com/facebookresearch/WSL-Images?\n\nThis model is pre-trained in a weakly supervised fashion on a dataset that is not available to the public (a bunch of instagram images), but is then finetuned on the public ImageNet1k like all the other common pre-trained models. \n\nWould this violate the rules?", "votes": 1, "replies": [ { "id": 641409, "postDate": "2019-10-04T17:00:11.330Z", "content": "

Please see this response I just made which should answer your question.

", "rawMarkdown": "Please see [this response](https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/110689#641407) I just made which should answer your question." } ] }, { "id": 636928, "postDate": "2019-09-30T11:24:30.233Z", "content": "

If you shared the pertained models you used for this contest in this discussion group then yes https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109274#latest-631802.

", "rawMarkdown": "If you shared the pertained models you used for this contest in this discussion group then yes https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109274#latest-631802.", "votes": 1, "replies": [ { "id": 638543, "postDate": "2019-10-02T02:49:36.060Z", "content": "

Thanks!

", "rawMarkdown": "Thanks!" } ] }, { "id": 637111, "postDate": "2019-09-30T17:05:19.070Z", "content": "

It sounds like you’re trying to use a pretrained model that was trained on a non-public dataset? If so, no, this is not permitted.

", "rawMarkdown": "It sounds like you’re trying to use a pretrained model that was trained on a non-public dataset? If so, no, this is not permitted.", "votes": -1, "replies": [ { "id": 638544, "postDate": "2019-10-02T02:49:57.073Z", "content": "

OK, thanks!

", "rawMarkdown": "OK, thanks!" }, { "id": 641195, "postDate": "2019-10-04T14:03:07.453Z", "content": "

@juliaelliott are you sure that's right? The pretrained model has publicly available weights. The only data that would be used in the competition is those pretrained weights - not the data that it was originally trained on. So AFAICT using pretrained models with publicly available weights is allowed according to the competition rules.

\n\n

If this really isn't the case, can I suggest you update the competition rules and publish a top-level thread saying so, otherwise people who haven't read down into this thread will never know and may end up disqualified over an issue they weren't aware of.

", "rawMarkdown": "@juliaelliott are you sure that's right? The pretrained model has publicly available weights. The only data that would be used in the competition is those pretrained weights - not the data that it was originally trained on. So AFAICT using pretrained models with publicly available weights is allowed according to the competition rules.\n\nIf this really isn't the case, can I suggest you update the competition rules and publish a top-level thread saying so, otherwise people who haven't read down into this thread will never know and may end up disqualified over an issue they weren't aware of.", "votes": 9 }, { "id": 641407, "postDate": "2019-10-04T16:58:02.920Z", "content": "

Thanks for your clarifications here. We agree that pretrained models on non-public datasets with publicly available weights are permitted.

", "rawMarkdown": "Thanks for your clarifications here. We agree that pretrained models on non-public datasets **with publicly available weights** are permitted.", "votes": 8 }, { "id": 647806, "postDate": "2019-10-13T10:01:10.917Z", "content": "

Thank you all for the clarification! I was just looking for this.

", "rawMarkdown": "Thank you all for the clarification! I was just looking for this.", "votes": 1 } ] } ], "comments": [ { "id": 640831, "author_name": "Tom Aindow", "author_url": "", "post_date": "2019-10-04T09:01:14.157000", "content": "

@juliaelliott Can you confirm if we are able to use https://github.com/facebookresearch/WSL-Images?

\n\n

This model is pre-trained in a weakly supervised fashion on a dataset that is not available to the public (a bunch of instagram images), but is then finetuned on the public ImageNet1k like all the other common pre-trained models.

\n\n

Would this violate the rules?

", "votes": 1, "replies": [ { "id": 641409, "author_name": "Julia Elliott", "author_url": "", "post_date": "2019-10-04T17:00:11.330000", "content": "

Please see this response I just made which should answer your question.

", "votes": 0, "replies": [] } ] }, { "id": 636928, "author_name": "SeshuRaju 🧘‍♂️", "author_url": "", "post_date": "2019-09-30T11:24:30.233000", "content": "

If you shared the pertained models you used for this contest in this discussion group then yes https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109274#latest-631802.

", "votes": 1, "replies": [ { "id": 638543, "author_name": "Mindy X", "author_url": "", "post_date": "2019-10-02T02:49:36.060000", "content": "

Thanks!

", "votes": 0, "replies": [] } ] }, { "id": 637111, "author_name": "Julia Elliott", "author_url": "", "post_date": "2019-09-30T17:05:19.070000", "content": "

It sounds like you’re trying to use a pretrained model that was trained on a non-public dataset? If so, no, this is not permitted.

", "votes": -1, "replies": [ { "id": 638544, "author_name": "Mindy X", "author_url": "", "post_date": "2019-10-02T02:49:57.073000", "content": "

OK, thanks!

", "votes": 0, "replies": [] }, { "id": 641195, "author_name": "Jeremy Howard", "author_url": "", "post_date": "2019-10-04T14:03:07.453000", "content": "

@juliaelliott are you sure that's right? The pretrained model has publicly available weights. The only data that would be used in the competition is those pretrained weights - not the data that it was originally trained on. So AFAICT using pretrained models with publicly available weights is allowed according to the competition rules.

\n\n

If this really isn't the case, can I suggest you update the competition rules and publish a top-level thread saying so, otherwise people who haven't read down into this thread will never know and may end up disqualified over an issue they weren't aware of.

", "votes": 9, "replies": [] }, { "id": 641407, "author_name": "Julia Elliott", "author_url": "", "post_date": "2019-10-04T16:58:02.920000", "content": "

Thanks for your clarifications here. We agree that pretrained models on non-public datasets with publicly available weights are permitted.

", "votes": 8, "replies": [] }, { "id": 647806, "author_name": "Carlo Lepelaars", "author_url": "", "post_date": "2019-10-13T10:01:10.917000", "content": "

Thank you all for the clarification! I was just looking for this.

", "votes": 1, "replies": [] } ] } ], "raw_markdown_by_id": { "636920": "Is it allowed to use publicly, freely available pretrained models trained on non-public dataset? Thanks!", "640831": "@juliaelliott Can you confirm if we are able to use https://github.com/facebookresearch/WSL-Images?\n\nThis model is pre-trained in a weakly supervised fashion on a dataset that is not available to the public (a bunch of instagram images), but is then finetuned on the public ImageNet1k like all the other common pre-trained models. \n\nWould this violate the rules?", "636928": "If you shared the pertained models you used for this contest in this discussion group then yes https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/discussion/109274#latest-631802.", "637111": "It sounds like you’re trying to use a pretrained model that was trained on a non-public dataset? If so, no, this is not permitted." } }