{ "id": 406810, "title": "Threshold for Segmentation", "url": "/competitions/vesuvius-challenge-ink-detection/discussion/406810", "author_name": "", "post_date": "2023-05-03T21:15:45.679002600Z", "votes": 1, "comment_count": 17, "views": 0, "content": "

What threshold are you guys trying out for segmentation? Are you guys adjusting the threshold according to the leaderboard score or setting it to a standard value?
\nI'm in a dilemma about the same.

", "messages": [ { "id": "2244736", "postDate": "05/03/2023 21:15:45", "content": "

What threshold are you guys trying out for segmentation? Are you guys adjusting the threshold according to the leaderboard score or setting it to a standard value?
\nI'm in a dilemma about the same.

", "rawMarkdown": "What threshold are you guys trying out for segmentation? Are you guys adjusting the threshold according to the leaderboard score or setting it to a standard value?\nI'm in a dilemma about the same.", "votes": null }, { "id": "2244891", "postDate": "05/04/2023 01:27:36", "content": "

I submit two models with different hyperparemeters :
\nOne for public leaderboard climbing which is probably very bad with a threshold lower than 0.5
\nAnd one of 0.5 which i assume to be better for the final leaderboard
\nI had the same dilemma but i am overwhelmed with the numbers of parameters so i stick with the strategy to avoid overthinking!!

", "rawMarkdown": "I submit two models with different hyperparemeters :\nOne for public leaderboard climbing which is probably very bad with a threshold lower than 0.5\nAnd one of 0.5 which i assume to be better for the final leaderboard\nI had the same dilemma but i am overwhelmed with the numbers of parameters so i stick with the strategy to avoid overthinking!!", "votes": null }, { "id": "2244942", "postDate": "05/04/2023 03:18:36", "content": "

How big of a gap do you see between your overfitted threshold and your .5 threshold?

", "rawMarkdown": "How big of a gap do you see between your overfitted threshold and your .5 threshold?", "votes": null }, { "id": "2245072", "postDate": "05/04/2023 06:13:43", "content": "

Yeah, I have done the same. There should be some other way to tackle this problem.

", "rawMarkdown": "Yeah, I have done the same. There should be some other way to tackle this problem.", "votes": null }, { "id": "2245139", "postDate": "05/04/2023 07:08:54", "content": "

I'm just setting my threshold to 0.5.

", "rawMarkdown": "I'm just setting my threshold to 0.5.", "votes": null }, { "id": "2245325", "postDate": "05/04/2023 09:30:01", "content": "

0.5-0.6 threshold am using…
\nfor how many epochs do you train?

", "rawMarkdown": "0.5-0.6 threshold am using...\nfor how many epochs do you train?", "votes": null }, { "id": "2245476", "postDate": "05/04/2023 11:59:23", "content": "

for the public model for +/- 0.1 treshold its +/- 0.08 score and +/- 0.1 between the two models something like this, in fact i prefer lie to myself thinking i have a well placement than face the reality that im totally out of topic :)

", "rawMarkdown": "for the public model for +/- 0.1 treshold its +/- 0.08 score and +/- 0.1 between the two models something like this, in fact i prefer lie to myself thinking i have a well placement than face the reality that im totally out of topic :)", "votes": null }, { "id": "2246860", "postDate": "05/05/2023 14:18:26", "content": "

I train for around 15 epochs.
\nHow long do you train? And what post-processing are you using?

", "rawMarkdown": "I train for around 15 epochs.\nHow long do you train? And what post-processing are you using?", "votes": null }, { "id": "2246862", "postDate": "05/05/2023 14:18:55", "content": "

Are you using any post-processing?

", "rawMarkdown": "Are you using any post-processing?", "votes": null }, { "id": "2247132", "postDate": "05/05/2023 18:29:30", "content": "

No. But I want to explore the post-processing methods on top in the nearest future.

", "rawMarkdown": "No. But I want to explore the post-processing methods on top in the nearest future.", "votes": null }, { "id": "2247672", "postDate": "05/06/2023 07:53:39", "content": "

12 epochs but slight overfitting ….

", "rawMarkdown": "12 epochs but slight overfitting ....", "votes": null }, { "id": "2248007", "postDate": "05/06/2023 12:43:06", "content": "

Are you ensembling models?
\nAlso, which loss are you using?

", "rawMarkdown": "Are you ensembling models? \nAlso, which loss are you using?", "votes": null }, { "id": "2248476", "postDate": "05/06/2023 22:04:56", "content": "

dice loss , 2 models ensembled

", "rawMarkdown": "dice loss , 2 models ensembled", "votes": null }, { "id": "2251498", "postDate": "05/09/2023 12:07:19", "content": "

Which losses and TTA are you using? I f you don't mind sharing

", "rawMarkdown": "Which losses and TTA are you using? I f you don't mind sharing", "votes": null }, { "id": "2251500", "postDate": "05/09/2023 12:07:44", "content": "

Are you using any TTA or post processing?

", "rawMarkdown": "Are you using any TTA or post processing?", "votes": null }, { "id": "2251561", "postDate": "05/09/2023 13:30:32", "content": "

tta: flips
\nloss: dice loss

", "rawMarkdown": "tta: flips \nloss: dice loss", "votes": null }, { "id": "2251891", "postDate": "05/09/2023 17:40:36", "content": "

Which TTA or post-processing are you using?

", "rawMarkdown": "Which TTA or post-processing are you using?", "votes": null }, { "id": "2257601", "postDate": "05/13/2023 14:03:27", "content": "

just dice loss？no BCE loss？

", "rawMarkdown": "just dice loss？no BCE loss？", "votes": null } ], "comments": [ { "id": 2244891, "author_name": "iraqbot", "author_url": "", "post_date": "05/04/2023 01:27:36", "content": "

I submit two models with different hyperparemeters :
\nOne for public leaderboard climbing which is probably very bad with a threshold lower than 0.5
\nAnd one of 0.5 which i assume to be better for the final leaderboard
\nI had the same dilemma but i am overwhelmed with the numbers of parameters so i stick with the strategy to avoid overthinking!!

", "votes": null, "replies": [ { "id": 2244942, "author_name": "petersk20", "author_url": "", "post_date": "05/04/2023 03:18:36", "content": "

How big of a gap do you see between your overfitted threshold and your .5 threshold?

", "votes": null, "replies": [ { "id": 2245476, "author_name": "iraqbot", "author_url": "", "post_date": "05/04/2023 11:59:23", "content": "

for the public model for +/- 0.1 treshold its +/- 0.08 score and +/- 0.1 between the two models something like this, in fact i prefer lie to myself thinking i have a well placement than face the reality that im totally out of topic :)

", "votes": null, "replies": [] } ] }, { "id": 2245072, "author_name": "datadixit", "author_url": "", "post_date": "05/04/2023 06:13:43", "content": "

Yeah, I have done the same. There should be some other way to tackle this problem.

", "votes": null, "replies": [] }, { "id": 2251891, "author_name": "datadixit", "author_url": "", "post_date": "05/09/2023 17:40:36", "content": "

Which TTA or post-processing are you using?

", "votes": null, "replies": [] } ] }, { "id": 2245139, "author_name": "igorkrashenyi", "author_url": "", "post_date": "05/04/2023 07:08:54", "content": "

I'm just setting my threshold to 0.5.

", "votes": null, "replies": [ { "id": 2246862, "author_name": "datadixit", "author_url": "", "post_date": "05/05/2023 14:18:55", "content": "

Are you using any post-processing?

", "votes": null, "replies": [ { "id": 2247132, "author_name": "igorkrashenyi", "author_url": "", "post_date": "05/05/2023 18:29:30", "content": "

No. But I want to explore the post-processing methods on top in the nearest future.

", "votes": null, "replies": [ { "id": 2251498, "author_name": "datadixit", "author_url": "", "post_date": "05/09/2023 12:07:19", "content": "

Which losses and TTA are you using? I f you don't mind sharing

", "votes": null, "replies": [ { "id": 2251561, "author_name": "igorkrashenyi", "author_url": "", "post_date": "05/09/2023 13:30:32", "content": "

tta: flips
\nloss: dice loss

", "votes": null, "replies": [] } ] } ] } ] } ] }, { "id": 2245325, "author_name": "arunodhayan", "author_url": "", "post_date": "05/04/2023 09:30:01", "content": "

0.5-0.6 threshold am using…
\nfor how many epochs do you train?

", "votes": null, "replies": [ { "id": 2246860, "author_name": "datadixit", "author_url": "", "post_date": "05/05/2023 14:18:26", "content": "

I train for around 15 epochs.
\nHow long do you train? And what post-processing are you using?

", "votes": null, "replies": [ { "id": 2247672, "author_name": "arunodhayan", "author_url": "", "post_date": "05/06/2023 07:53:39", "content": "

12 epochs but slight overfitting ….

", "votes": null, "replies": [ { "id": 2248007, "author_name": "datadixit", "author_url": "", "post_date": "05/06/2023 12:43:06", "content": "

Are you ensembling models?
\nAlso, which loss are you using?

", "votes": null, "replies": [ { "id": 2248476, "author_name": "arunodhayan", "author_url": "", "post_date": "05/06/2023 22:04:56", "content": "

dice loss , 2 models ensembled

", "votes": null, "replies": [ { "id": 2251500, "author_name": "datadixit", "author_url": "", "post_date": "05/09/2023 12:07:44", "content": "

Are you using any TTA or post processing?

", "votes": null, "replies": [] }, { "id": 2257601, "author_name": "wangxuc", "author_url": "", "post_date": "05/13/2023 14:03:27", "content": "

just dice loss？no BCE loss？

", "votes": null, "replies": [] } ] } ] } ] } ] } ] } ], "raw_markdown_by_id": { "2244736": "What threshold are you guys trying out for segmentation? Are you guys adjusting the threshold according to the leaderboard score or setting it to a standard value?\nI'm in a dilemma about the same.", "2244891": "I submit two models with different hyperparemeters :\nOne for public leaderboard climbing which is probably very bad with a threshold lower than 0.5\nAnd one of 0.5 which i assume to be better for the final leaderboard\nI had the same dilemma but i am overwhelmed with the numbers of parameters so i stick with the strategy to avoid overthinking!!", "2244942": "How big of a gap do you see between your overfitted threshold and your .5 threshold?", "2245072": "Yeah, I have done the same. There should be some other way to tackle this problem.", "2245139": "I'm just setting my threshold to 0.5.", "2245325": "0.5-0.6 threshold am using...\nfor how many epochs do you train?", "2245476": "for the public model for +/- 0.1 treshold its +/- 0.08 score and +/- 0.1 between the two models something like this, in fact i prefer lie to myself thinking i have a well placement than face the reality that im totally out of topic :)", "2246860": "I train for around 15 epochs.\nHow long do you train? And what post-processing are you using?", "2246862": "Are you using any post-processing?", "2247132": "No. But I want to explore the post-processing methods on top in the nearest future.", "2247672": "12 epochs but slight overfitting ....", "2248007": "Are you ensembling models? \nAlso, which loss are you using?", "2248476": "dice loss , 2 models ensembled", "2251498": "Which losses and TTA are you using? I f you don't mind sharing", "2251500": "Are you using any TTA or post processing?", "2251561": "tta: flips \nloss: dice loss", "2251891": "Which TTA or post-processing are you using?", "2257601": "just dice loss？no BCE loss？" }, "source": "meta" }