Hello all! I have not competed in many Kaggle competitions, and I have a question about approach:
\nSo far I have been doing training + inference in my notebooks for submission. However, based on some threads it seems like others might be loading a model they have trained locally/elsewhere, and just doing inference during submission time.
\nIt seems like it is in the spirit of the competition to submit your code with your model, but that does put an upper cap on the complexity of the models we can submit. Would it disqualify me to do this? The FAQ and submission guidelines are not clear on this.
", "messages": [ { "id": "3333902", "postDate": "11/17/2025 10:43:04", "content": "Hello all! I have not competed in many Kaggle competitions, and I have a question about approach:
\nSo far I have been doing training + inference in my notebooks for submission. However, based on some threads it seems like others might be loading a model they have trained locally/elsewhere, and just doing inference during submission time.
\nIt seems like it is in the spirit of the competition to submit your code with your model, but that does put an upper cap on the complexity of the models we can submit. Would it disqualify me to do this? The FAQ and submission guidelines are not clear on this.
", "rawMarkdown": "Hello all! I have not competed in many Kaggle competitions, and I have a question about approach:\n\nSo far I have been doing training + inference in my notebooks for submission. However, based on some threads it seems like others might be loading a model they have trained locally/elsewhere, and just doing inference during submission time.\n\nIt seems like it is in the spirit of the competition to submit your code with your model, but that does put an upper cap on the complexity of the models we can submit. Would it disqualify me to do this? The FAQ and submission guidelines are not clear on this.", "votes": null }, { "id": "3341190", "postDate": "11/20/2025 03:17:05", "content": "I don't think this should disqualify you, however the training of the model should be replicable for safety reasons.
", "rawMarkdown": "I don't think this should disqualify you, however the training of the model should be replicable for safety reasons.", "votes": null }, { "id": "3341365", "postDate": "11/20/2025 06:05:26", "content": "It depends on what your goal is. For example, if you want to prove that your training delivers a certain score and be transparent, then you want to have training and inference in a single notebook, as I did here: https://www.kaggle.com/code/jirkaborovec/surface-detect-baseline-slice-2d-segmentation\nIf your training takes too long or you need a more capable HW, then you train the model elsewhere and just load the model for inference. The idea is that you do not use your limited GPU time given by Kaggle, as submission does not count into this time.
", "rawMarkdown": "It depends on what your goal is. For example, if you want to prove that your training delivers a certain score and be transparent, then you want to have training and inference in a single notebook, as I did here: https://www.kaggle.com/code/jirkaborovec/surface-detect-baseline-slice-2d-segmentation\nIf your training takes too long or you need a more capable HW, then you train the model elsewhere and just load the model for inference. The idea is that you do not use your limited GPU time given by Kaggle, as submission does not count into this time.", "votes": null } ], "comments": [ { "id": 3341190, "author_name": "giorgioangelotti", "author_url": "", "post_date": "11/20/2025 03:17:05", "content": "I don't think this should disqualify you, however the training of the model should be replicable for safety reasons.
", "votes": null, "replies": [] }, { "id": 3341365, "author_name": "jirkaborovec", "author_url": "", "post_date": "11/20/2025 06:05:26", "content": "It depends on what your goal is. For example, if you want to prove that your training delivers a certain score and be transparent, then you want to have training and inference in a single notebook, as I did here: https://www.kaggle.com/code/jirkaborovec/surface-detect-baseline-slice-2d-segmentation\nIf your training takes too long or you need a more capable HW, then you train the model elsewhere and just load the model for inference. The idea is that you do not use your limited GPU time given by Kaggle, as submission does not count into this time.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3333902": "Hello all! I have not competed in many Kaggle competitions, and I have a question about approach:\n\nSo far I have been doing training + inference in my notebooks for submission. However, based on some threads it seems like others might be loading a model they have trained locally/elsewhere, and just doing inference during submission time.\n\nIt seems like it is in the spirit of the competition to submit your code with your model, but that does put an upper cap on the complexity of the models we can submit. Would it disqualify me to do this? The FAQ and submission guidelines are not clear on this.", "3341190": "I don't think this should disqualify you, however the training of the model should be replicable for safety reasons.", "3341365": "It depends on what your goal is. For example, if you want to prove that your training delivers a certain score and be transparent, then you want to have training and inference in a single notebook, as I did here: https://www.kaggle.com/code/jirkaborovec/surface-detect-baseline-slice-2d-segmentation\nIf your training takes too long or you need a more capable HW, then you train the model elsewhere and just load the model for inference. The idea is that you do not use your limited GPU time given by Kaggle, as submission does not count into this time." }, "source": "meta" }