{ "id": 295846, "title": "Why the MIN_PIXELS = {0: 75, 1:75, 2:150} ? Has anyone done experiment on it?", "url": "/competitions/sartorius-cell-instance-segmentation/discussion/295846", "author_name": "", "post_date": "2021-12-18T09:18:52.264381500Z", "votes": 2, "comment_count": 7, "views": 0, "content": "

In infer stage, we set the mask.sum()>MIN_PIXELS[pred_cls]. About optimal threshold, has anyone done experiment on it?

", "messages": [ { "id": "1621978", "postDate": "12/18/2021 09:18:52", "content": "

In infer stage, we set the mask.sum()>MIN_PIXELS[pred_cls]. About optimal threshold, has anyone done experiment on it?

", "rawMarkdown": "In infer stage, we set the mask.sum()>MIN_PIXELS[pred_cls]. About optimal threshold, has anyone done experiment on it?", "votes": null }, { "id": "1622219", "postDate": "12/18/2021 13:47:00", "content": "

Different threshold have different optimal MIN_PIXELS.

\n

An optimal threshold for a single fold CV maybe a overfitting result of validation dataset.

", "rawMarkdown": "Different threshold have different optimal MIN_PIXELS.\n\nAn optimal threshold for a single fold CV maybe a overfitting result of validation dataset.", "votes": null }, { "id": "1622421", "postDate": "12/18/2021 18:02:28", "content": "

I find my own post-processing thresholds with cross-validation. You shouldn't trust the ones that are publicly shared.

", "rawMarkdown": "I find my own post-processing thresholds with cross-validation. You shouldn't trust the ones that are publicly shared.", "votes": null }, { "id": "1622625", "postDate": "12/19/2021 01:10:49", "content": "

Hello, what cross-validation method do you use? GroupKFold or sth?

", "rawMarkdown": "Hello, what cross-validation method do you use? GroupKFold or sth?", "votes": null }, { "id": "1622667", "postDate": "12/19/2021 03:03:14", "content": "
\n

I find my own post-processing thresholds with cross-validation.

\n
\n

May I ask are you averaging the optimal thresholds across all folds for the final threshold for inference?

", "rawMarkdown": "> I find my own post-processing thresholds with cross-validation.\n\nMay I ask are you averaging the optimal thresholds across all folds for the final threshold for inference?", "votes": null }, { "id": "1622720", "postDate": "12/19/2021 04:35:44", "content": "

I use stratified folds mentioned here, and I use thresholds that maximizes the OOF score.

", "rawMarkdown": "I use stratified folds mentioned [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/285546), and I use thresholds that maximizes the OOF score.", "votes": null }, { "id": "1622734", "postDate": "12/19/2021 05:09:49", "content": "

Thank you for your reply.

", "rawMarkdown": "Thank you for your reply.", "votes": null }, { "id": "1622736", "postDate": "12/19/2021 05:15:11", "content": "

Thanks!

", "rawMarkdown": "Thanks!", "votes": null } ], "comments": [ { "id": 1622219, "author_name": "blueboy97", "author_url": "", "post_date": "12/18/2021 13:47:00", "content": "

Different threshold have different optimal MIN_PIXELS.

\n

An optimal threshold for a single fold CV maybe a overfitting result of validation dataset.

", "votes": null, "replies": [] }, { "id": 1622421, "author_name": "gunesevitan", "author_url": "", "post_date": "12/18/2021 18:02:28", "content": "

I find my own post-processing thresholds with cross-validation. You shouldn't trust the ones that are publicly shared.

", "votes": null, "replies": [ { "id": 1622625, "author_name": "blueboy97", "author_url": "", "post_date": "12/19/2021 01:10:49", "content": "

Hello, what cross-validation method do you use? GroupKFold or sth?

", "votes": null, "replies": [] }, { "id": 1622667, "author_name": "atharvaingle", "author_url": "", "post_date": "12/19/2021 03:03:14", "content": "
\n

I find my own post-processing thresholds with cross-validation.

\n
\n

May I ask are you averaging the optimal thresholds across all folds for the final threshold for inference?

", "votes": null, "replies": [] }, { "id": 1622720, "author_name": "gunesevitan", "author_url": "", "post_date": "12/19/2021 04:35:44", "content": "

I use stratified folds mentioned here, and I use thresholds that maximizes the OOF score.

", "votes": null, "replies": [] }, { "id": 1622734, "author_name": "blueboy97", "author_url": "", "post_date": "12/19/2021 05:09:49", "content": "

Thank you for your reply.

", "votes": null, "replies": [] }, { "id": 1622736, "author_name": "deeeeeeeplearning", "author_url": "", "post_date": "12/19/2021 05:15:11", "content": "

Thanks!

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1621978": "In infer stage, we set the mask.sum()>MIN_PIXELS[pred_cls]. About optimal threshold, has anyone done experiment on it?", "1622219": "Different threshold have different optimal MIN_PIXELS.\n\nAn optimal threshold for a single fold CV maybe a overfitting result of validation dataset.", "1622421": "I find my own post-processing thresholds with cross-validation. You shouldn't trust the ones that are publicly shared.", "1622625": "Hello, what cross-validation method do you use? GroupKFold or sth?", "1622667": "> I find my own post-processing thresholds with cross-validation.\n\nMay I ask are you averaging the optimal thresholds across all folds for the final threshold for inference?", "1622720": "I use stratified folds mentioned [here](https://www.kaggle.com/c/sartorius-cell-instance-segmentation/discussion/285546), and I use thresholds that maximizes the OOF score.", "1622734": "Thank you for your reply.", "1622736": "Thanks!" }, "source": "meta" }