Hi everyone,
\nI’d like to clarify one point about the rules. Since the test image files in the submitting environment (without labels) can be got, is it considered rule-compliant to do self-supervised / unsupervised pretraining (e.g., MAE/DINO/SimCLR) on the test images and then fine-tune only on the labeled training set? Because the training dataset is quite small, I suspect that doing unsupervised (self-supervised) learning on the test set may help the model achieve better performance. However, it might not align with real-world generalization requirements, so I’m asking everyone here for clarification/opinions.
\nThanks in advance!
Hi everyone,
\nI’d like to clarify one point about the rules. Since the test image files in the submitting environment (without labels) can be got, is it considered rule-compliant to do self-supervised / unsupervised pretraining (e.g., MAE/DINO/SimCLR) on the test images and then fine-tune only on the labeled training set? Because the training dataset is quite small, I suspect that doing unsupervised (self-supervised) learning on the test set may help the model achieve better performance. However, it might not align with real-world generalization requirements, so I’m asking everyone here for clarification/opinions.
\nThanks in advance!
I do not see a problem with this approach. It should be allowed
", "rawMarkdown": "I do not see a problem with this approach. It should be allowed", "votes": null }, { "id": "3381932", "postDate": "12/26/2025 00:37:35", "content": "basically pseudo labelling sort of thing, which works most of the time and there are no rules against it. however, the test data in this competition specifically is private and can't be accessed by anyone.
", "rawMarkdown": "basically pseudo labelling sort of thing, which works most of the time and there are no rules against it. however, the test data in this competition specifically is private and can't be accessed by anyone.", "votes": null }, { "id": "3381935", "postDate": "12/26/2025 02:01:07", "content": "Your incorrect about data access.
\nWhile we don't get access to the data to see it, the final notebook can certainly train your model using student - teacher type methods as it sees the data to make the predictions. Assuming there are 800 images you might want to up sample training images to 3*357 to fine tune with your 800 test \"predictions\".
\nThe major drawback is time - you probably have 5 folds to your model. You need to tune all 5 and also predict. My initial attempt a few weeks ago ran over the 9 hour compute limit on a pretty small model. So not sure it will work as I got no LB score on the attempt. If you best work is an ensemble of 3 models than my guess you will run over the 9 hour limit with anything that might have generated a score improvement.
\nWith this particular data set it should give you a boost to a model if you can do it in less than 9 hours. When it worked for me in the past I needed about 5 epochs of training time. I assume it's ok to the rules but don't know that for sure as I was never high enough on the metals LB to have my work checked.
", "rawMarkdown": "Your incorrect about data access. \n\nWhile we don't get access to the data to see it, the final notebook can certainly train your model using student - teacher type methods as it sees the data to make the predictions. Assuming there are 800 images you might want to up sample training images to 3*357 to fine tune with your 800 test \"predictions\".\n\nThe major drawback is time - you probably have 5 folds to your model. You need to tune all 5 and also predict. My initial attempt a few weeks ago ran over the 9 hour compute limit on a pretty small model. So not sure it will work as I got no LB score on the attempt. If you best work is an ensemble of 3 models than my guess you will run over the 9 hour limit with anything that might have generated a score improvement.\n\nWith this particular data set it should give you a boost to a model if you can do it in less than 9 hours. When it worked for me in the past I needed about 5 epochs of training time. I assume it's ok to the rules but don't know that for sure as I was never high enough on the metals LB to have my work checked.", "votes": null }, { "id": "3381946", "postDate": "12/26/2025 02:28:27", "content": "Thank you!
", "rawMarkdown": "Thank you!", "votes": null }, { "id": "3381947", "postDate": "12/26/2025 02:30:40", "content": "Thank you for your detailed reply! Your experience looks very valuable.
", "rawMarkdown": "Thank you for your detailed reply! Your experience looks very valuable.", "votes": null }, { "id": "3382702", "postDate": "12/28/2025 10:07:11", "content": "كتير من مشكل
", "rawMarkdown": "كتير من مشكل", "votes": null } ], "comments": [ { "id": 3381879, "author_name": "jeannkouagou", "author_url": "", "post_date": "12/25/2025 19:42:59", "content": "I do not see a problem with this approach. It should be allowed
", "votes": null, "replies": [ { "id": 3381946, "author_name": "co000l", "author_url": "", "post_date": "12/26/2025 02:28:27", "content": "Thank you!
", "votes": null, "replies": [] } ] }, { "id": 3381932, "author_name": "amjedalqasemi", "author_url": "", "post_date": "12/26/2025 00:37:35", "content": "basically pseudo labelling sort of thing, which works most of the time and there are no rules against it. however, the test data in this competition specifically is private and can't be accessed by anyone.
", "votes": null, "replies": [ { "id": 3381935, "author_name": "pcjimmmy", "author_url": "", "post_date": "12/26/2025 02:01:07", "content": "Your incorrect about data access.
\nWhile we don't get access to the data to see it, the final notebook can certainly train your model using student - teacher type methods as it sees the data to make the predictions. Assuming there are 800 images you might want to up sample training images to 3*357 to fine tune with your 800 test \"predictions\".
\nThe major drawback is time - you probably have 5 folds to your model. You need to tune all 5 and also predict. My initial attempt a few weeks ago ran over the 9 hour compute limit on a pretty small model. So not sure it will work as I got no LB score on the attempt. If you best work is an ensemble of 3 models than my guess you will run over the 9 hour limit with anything that might have generated a score improvement.
\nWith this particular data set it should give you a boost to a model if you can do it in less than 9 hours. When it worked for me in the past I needed about 5 epochs of training time. I assume it's ok to the rules but don't know that for sure as I was never high enough on the metals LB to have my work checked.
", "votes": null, "replies": [ { "id": 3381947, "author_name": "co000l", "author_url": "", "post_date": "12/26/2025 02:30:40", "content": "Thank you for your detailed reply! Your experience looks very valuable.
", "votes": null, "replies": [] } ] } ] }, { "id": 3382702, "author_name": "khalifafaris", "author_url": "", "post_date": "12/28/2025 10:07:11", "content": "كتير من مشكل
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3381310": "Hi everyone, \nI’d like to clarify one point about the rules. Since the test image files in the submitting environment (without labels) can be got, is it considered rule-compliant to do self-supervised / unsupervised pretraining (e.g., MAE/DINO/SimCLR) on the test images and then fine-tune only on the labeled training set? Because the training dataset is quite small, I suspect that doing unsupervised (self-supervised) learning on the test set may help the model achieve better performance. However, it might not align with real-world generalization requirements, so I’m asking everyone here for clarification/opinions. \nThanks in advance!", "3381879": "I do not see a problem with this approach. It should be allowed", "3381932": "basically pseudo labelling sort of thing, which works most of the time and there are no rules against it. however, the test data in this competition specifically is private and can't be accessed by anyone.", "3381935": "Your incorrect about data access. \n\nWhile we don't get access to the data to see it, the final notebook can certainly train your model using student - teacher type methods as it sees the data to make the predictions. Assuming there are 800 images you might want to up sample training images to 3*357 to fine tune with your 800 test \"predictions\".\n\nThe major drawback is time - you probably have 5 folds to your model. You need to tune all 5 and also predict. My initial attempt a few weeks ago ran over the 9 hour compute limit on a pretty small model. So not sure it will work as I got no LB score on the attempt. If you best work is an ensemble of 3 models than my guess you will run over the 9 hour limit with anything that might have generated a score improvement.\n\nWith this particular data set it should give you a boost to a model if you can do it in less than 9 hours. When it worked for me in the past I needed about 5 epochs of training time. I assume it's ok to the rules but don't know that for sure as I was never high enough on the metals LB to have my work checked.", "3381946": "Thank you!", "3381947": "Thank you for your detailed reply! Your experience looks very valuable.", "3382702": "كتير من مشكل" }, "source": "meta" }