{ "id": 234647, "title": "Code vs Kernel competition training clarification", "url": "/competitions/birdclef-2021/discussion/234647", "author_name": "", "post_date": "2021-04-25T11:51:22.694310300Z", "votes": 2, "comment_count": 3, "views": 0, "content": "

I'm pretty new to competitions here, so wanted to be clear about something before I got started - sorry if I'm missing something super obvious!

\n

My question is about when and how you are allowed to train models in this competition vs other code competitions:

\n

In the Code Competition FAQs (https://www.kaggle.com/docs/competitions) it links to the Quora competition (https://www.kaggle.com/c/quora-insincere-questions-classification) as an example of a code competition - but the Quora competition is also a \"kernel\" competition.

\n

In the Quora competition, the Kernel-FAQ says that both training and prediction have to be done in a single kernel, under the GPU limits.

\n

This competition (Birdcall) has no such requirement though, is that correct? Is that because it is a \"code\" competition, but not a \"kernel\" competition? Am I missing somewhere where that rule is explicitly called out, or should I always assume it is ok if it doesn't have the same wording as that Quora competition?

\n

Thanks, and sorry again if this is obvious and I missed it!

\n

(This same basic question was asked in the Coleridge Initiative competition, but I felt the need to check for this one as well: https://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/230618)

", "messages": [ { "id": "1283913", "postDate": "04/25/2021 11:51:22", "content": "

I'm pretty new to competitions here, so wanted to be clear about something before I got started - sorry if I'm missing something super obvious!

\n

My question is about when and how you are allowed to train models in this competition vs other code competitions:

\n

In the Code Competition FAQs (https://www.kaggle.com/docs/competitions) it links to the Quora competition (https://www.kaggle.com/c/quora-insincere-questions-classification) as an example of a code competition - but the Quora competition is also a \"kernel\" competition.

\n

In the Quora competition, the Kernel-FAQ says that both training and prediction have to be done in a single kernel, under the GPU limits.

\n

This competition (Birdcall) has no such requirement though, is that correct? Is that because it is a \"code\" competition, but not a \"kernel\" competition? Am I missing somewhere where that rule is explicitly called out, or should I always assume it is ok if it doesn't have the same wording as that Quora competition?

\n

Thanks, and sorry again if this is obvious and I missed it!

\n

(This same basic question was asked in the Coleridge Initiative competition, but I felt the need to check for this one as well: https://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/230618)

", "rawMarkdown": "I'm pretty new to competitions here, so wanted to be clear about something before I got started - sorry if I'm missing something super obvious!\n\nMy question is about when and how you are allowed to train models in this competition vs other code competitions:\n\nIn the Code Competition FAQs (https://www.kaggle.com/docs/competitions) it links to the Quora competition (https://www.kaggle.com/c/quora-insincere-questions-classification) as an example of a code competition - but the Quora competition is also a \"kernel\" competition.\n\nIn the Quora competition, the Kernel-FAQ says that both training and prediction have to be done in a single kernel, under the GPU limits.\n\nThis competition (Birdcall) has no such requirement though, is that correct? Is that because it is a \"code\" competition, but not a \"kernel\" competition? Am I missing somewhere where that rule is explicitly called out, or should I always assume it is ok if it doesn't have the same wording as that Quora competition?\n\nThanks, and sorry again if this is obvious and I missed it!\n\n(This same basic question was asked in the Coleridge Initiative competition, but I felt the need to check for this one as well: https://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/230618)", "votes": null }, { "id": "1284202", "postDate": "04/25/2021 17:15:21", "content": "

EDIT: in this thread https://www.kaggle.com/c/birdclef-2021/discussion/233539 Stefan makes it clear that it's ok to use private pre-trained models that you train yourself, but I guess my question is if that is just true for all code competitions unless otherwise stated?

\n

I wouldn't even have a question about it (assuming that pre-training and using your own model weights would be obviously fine), except in light of the Quora kernel competition which said that you couldn't, and this line in this competition's rules:

\n
\n

Freely & publicly available external data is allowed, including pre-trained models

\n
\n

which seems to suggest that you could only use pre-trained models (including your own) if they are made publicly available.

\n

Again - sorry if I just totally missed something! But it does seem that several people had the same general question, since I've now found multiple threads about it :)

", "rawMarkdown": "EDIT: in this thread https://www.kaggle.com/c/birdclef-2021/discussion/233539 Stefan makes it clear that it's ok to use private pre-trained models that you train yourself, but I guess my question is if that is just true for all code competitions unless otherwise stated?\n\nI wouldn't even have a question about it (assuming that pre-training and using your own model weights would be obviously fine), except in light of the Quora kernel competition which said that you couldn't, and this line in this competition's rules:\n\n> Freely & publicly available external data is allowed, including pre-trained models\n\nwhich seems to suggest that you could only use pre-trained models (including your own) if they are made publicly available.\n\nAgain - sorry if I just totally missed something! But it does seem that several people had the same general question, since I've now found multiple threads about it :)", "votes": null }, { "id": "1284590", "postDate": "04/26/2021 05:54:08", "content": "

Code competitions are the same as what was called kernel competitions.

\n

Code competitions means that at least inference has to be made in a notebook or a script without internet access. Then code competitions each have their own requirements, like compute limit (max duration), or need to also train models in notebook. In the latter case train data is not fully available outside submitted notebooks.

\n

Here, the code tab under the overview menu does not require training to be done in notebook.

", "rawMarkdown": "Code competitions are the same as what was called kernel competitions.\n\nCode competitions means that at least inference has to be made in a notebook or a script without internet access. Then code competitions each have their own requirements, like compute limit (max duration), or need to also train models in notebook. In the latter case train data is not fully available outside submitted notebooks.\n\nHere, the code tab under the overview menu does not require training to be done in notebook.", "votes": null }, { "id": "1284753", "postDate": "04/26/2021 08:41:37", "content": "

Ahh, got it - thank you!

", "rawMarkdown": "Ahh, got it - thank you!", "votes": null } ], "comments": [ { "id": 1284202, "author_name": "chris62", "author_url": "", "post_date": "04/25/2021 17:15:21", "content": "

EDIT: in this thread https://www.kaggle.com/c/birdclef-2021/discussion/233539 Stefan makes it clear that it's ok to use private pre-trained models that you train yourself, but I guess my question is if that is just true for all code competitions unless otherwise stated?

\n

I wouldn't even have a question about it (assuming that pre-training and using your own model weights would be obviously fine), except in light of the Quora kernel competition which said that you couldn't, and this line in this competition's rules:

\n
\n

Freely & publicly available external data is allowed, including pre-trained models

\n
\n

which seems to suggest that you could only use pre-trained models (including your own) if they are made publicly available.

\n

Again - sorry if I just totally missed something! But it does seem that several people had the same general question, since I've now found multiple threads about it :)

", "votes": null, "replies": [ { "id": 1284590, "author_name": "cpmpml", "author_url": "", "post_date": "04/26/2021 05:54:08", "content": "

Code competitions are the same as what was called kernel competitions.

\n

Code competitions means that at least inference has to be made in a notebook or a script without internet access. Then code competitions each have their own requirements, like compute limit (max duration), or need to also train models in notebook. In the latter case train data is not fully available outside submitted notebooks.

\n

Here, the code tab under the overview menu does not require training to be done in notebook.

", "votes": null, "replies": [] }, { "id": 1284753, "author_name": "chris62", "author_url": "", "post_date": "04/26/2021 08:41:37", "content": "

Ahh, got it - thank you!

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1283913": "I'm pretty new to competitions here, so wanted to be clear about something before I got started - sorry if I'm missing something super obvious!\n\nMy question is about when and how you are allowed to train models in this competition vs other code competitions:\n\nIn the Code Competition FAQs (https://www.kaggle.com/docs/competitions) it links to the Quora competition (https://www.kaggle.com/c/quora-insincere-questions-classification) as an example of a code competition - but the Quora competition is also a \"kernel\" competition.\n\nIn the Quora competition, the Kernel-FAQ says that both training and prediction have to be done in a single kernel, under the GPU limits.\n\nThis competition (Birdcall) has no such requirement though, is that correct? Is that because it is a \"code\" competition, but not a \"kernel\" competition? Am I missing somewhere where that rule is explicitly called out, or should I always assume it is ok if it doesn't have the same wording as that Quora competition?\n\nThanks, and sorry again if this is obvious and I missed it!\n\n(This same basic question was asked in the Coleridge Initiative competition, but I felt the need to check for this one as well: https://www.kaggle.com/c/coleridgeinitiative-show-us-the-data/discussion/230618)", "1284202": "EDIT: in this thread https://www.kaggle.com/c/birdclef-2021/discussion/233539 Stefan makes it clear that it's ok to use private pre-trained models that you train yourself, but I guess my question is if that is just true for all code competitions unless otherwise stated?\n\nI wouldn't even have a question about it (assuming that pre-training and using your own model weights would be obviously fine), except in light of the Quora kernel competition which said that you couldn't, and this line in this competition's rules:\n\n> Freely & publicly available external data is allowed, including pre-trained models\n\nwhich seems to suggest that you could only use pre-trained models (including your own) if they are made publicly available.\n\nAgain - sorry if I just totally missed something! But it does seem that several people had the same general question, since I've now found multiple threads about it :)", "1284590": "Code competitions are the same as what was called kernel competitions.\n\nCode competitions means that at least inference has to be made in a notebook or a script without internet access. Then code competitions each have their own requirements, like compute limit (max duration), or need to also train models in notebook. In the latter case train data is not fully available outside submitted notebooks.\n\nHere, the code tab under the overview menu does not require training to be done in notebook.", "1284753": "Ahh, got it - thank you!" }, "source": "meta" }