{ "id": 348888, "title": "What's your HuBMAP only score?", "url": "/competitions/hubmap-organ-segmentation/discussion/348888", "author_name": "", "post_date": "2022-08-30T13:04:50.029767Z", "votes": 17, "comment_count": 38, "views": 0, "content": "

I decided to submit only HuBMAP after reaching 0.8 since I want to use my last submissions efficiently. My best submission scores 0.58 on HuBMAP only part of the test set.

\n

Edit: I just reached 0.59. I guess 0.62 is the gold zone.
\nEdit 2: I overestimated the gold zone. I'm still struggling to reach 0.60.

", "messages": [ { "id": "1919500", "postDate": "08/30/2022 13:04:50", "content": "

I decided to submit only HuBMAP after reaching 0.8 since I want to use my last submissions efficiently. My best submission scores 0.58 on HuBMAP only part of the test set.

\n

Edit: I just reached 0.59. I guess 0.62 is the gold zone.
\nEdit 2: I overestimated the gold zone. I'm still struggling to reach 0.60.

", "rawMarkdown": "I decided to submit only HuBMAP after reaching 0.8 since I want to use my last submissions efficiently. My best submission scores 0.58 on HuBMAP only part of the test set.\n\n**Edit**: I just reached 0.59. I guess 0.62 is the gold zone.\n**Edit 2**: I overestimated the gold zone. I'm still struggling to reach 0.60.", "votes": null }, { "id": "1919549", "postDate": "08/30/2022 14:00:33", "content": "

Thanks for your sharing! Could you tell me how to know the my submission scores on HuBMAP?

", "rawMarkdown": "Thanks for your sharing! Could you tell me how to know the my submission scores on HuBMAP?", "votes": null }, { "id": "1919568", "postDate": "08/30/2022 14:27:51", "content": "

Assign empty string to HPA samples.

", "rawMarkdown": "Assign empty string to HPA samples.", "votes": null }, { "id": "1919688", "postDate": "08/30/2022 15:54:08", "content": "

same here 0.58 🤓

", "rawMarkdown": "same here 0.58 🤓", "votes": null }, { "id": "1919754", "postDate": "08/30/2022 16:47:18", "content": "

check this out https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333631#1835768

", "rawMarkdown": "check this out https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333631#1835768", "votes": null }, { "id": "1920752", "postDate": "08/31/2022 11:36:41", "content": "

My best submission scores 0.60 on HuBMAP，it's really hard to improve the performance on HuBMAP😂

", "rawMarkdown": "My best submission scores 0.60 on HuBMAP，it's really hard to improve the performance on HuBMAP😂", "votes": null }, { "id": "1920899", "postDate": "08/31/2022 13:11:36", "content": "

Amazing👍👍

", "rawMarkdown": "Amazing👍👍", "votes": null }, { "id": "1921323", "postDate": "08/31/2022 18:02:42", "content": "

0.57 using 4 folds of deeplabV3-efficientnetB4.
\nIs your 0.58 and 0.59 ensembling, n-folds, or single model?

", "rawMarkdown": "0.57 using 4 folds of deeplabV3-efficientnetB4.\nIs your 0.58 and 0.59 ensembling, n-folds, or single model?", "votes": null }, { "id": "1921343", "postDate": "08/31/2022 18:42:23", "content": "

5 folds of single model.

", "rawMarkdown": "5 folds of single model.", "votes": null }, { "id": "1921793", "postDate": "09/01/2022 04:20:23", "content": "

Do you mean single model performs better?

", "rawMarkdown": "Do you mean single model performs better?", "votes": null }, { "id": "1921903", "postDate": "09/01/2022 06:18:07", "content": "

I mean I'm using average of 5 folds of a model. Most of the time, that's what we refer to as a single model on Kaggle. I haven't started my ensemble yet. I'm still thinking about what kind of diversity should I bring with ensemble.

", "rawMarkdown": "I mean I'm using average of 5 folds of a model. Most of the time, that's what we refer to as a single model on Kaggle. I haven't started my ensemble yet. I'm still thinking about what kind of diversity should I bring with ensemble.", "votes": null }, { "id": "1922018", "postDate": "09/01/2022 08:04:13", "content": "

same as 0.59, 5fold of single model.

", "rawMarkdown": "same as 0.59, 5fold of single model.", "votes": null }, { "id": "1922164", "postDate": "09/01/2022 10:06:02", "content": "

Why did my hubmap score stuck in 0.49? Is there anybody willing to help me find the problem?

", "rawMarkdown": "Why did my hubmap score stuck in 0.49? Is there anybody willing to help me find the problem?", "votes": null }, { "id": "1922182", "postDate": "09/01/2022 10:20:41", "content": "

Can you show your predictions on the single test image?

", "rawMarkdown": "Can you show your predictions on the single test image?", "votes": null }, { "id": "1922219", "postDate": "09/01/2022 10:55:58", "content": "

same as 0.59☝️

", "rawMarkdown": "same as 0.59☝️", "votes": null }, { "id": "1922250", "postDate": "09/01/2022 11:27:38", "content": "

\"\"
\nDuring my local validation, in fold0, local cv reach 0.8269, which seems a good improvement. But lb only gets 0.70, and hubmap score is 0.49. It seems that my test results have many false positives.

", "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6149314%2F7e296638ef80efff944d0b797aeea0a6%2F10078_050.jpg?generation=1662031531847507&alt=media)\nDuring my local validation, in fold0, local cv reach 0.8269, which seems a good improvement. But lb only gets 0.70, and hubmap score is 0.49. It seems that my test results have many false positives.", "votes": null }, { "id": "1922376", "postDate": "09/01/2022 12:50:34", "content": "

Maybe you can try to adjust the mask threshold in test images.

", "rawMarkdown": "Maybe you can try to adjust the mask threshold in test images.", "votes": null }, { "id": "1924811", "postDate": "09/03/2022 12:34:34", "content": "

\"\"
\nMy new test result was presented above. Many false positives has been removed, and my hubmap score has achieved 0.52. Could anybody help me to further improve my score?

", "rawMarkdown": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6149314%2F786878a0b0ee82545e519449d4cecfdb%2F10078_092_b5_040.jpg?generation=1662208416757805&alt=media)\nMy new test result was presented above. Many false positives has been removed, and my hubmap score has achieved 0.52. Could anybody help me to further improve my score?", "votes": null }, { "id": "1928275", "postDate": "09/06/2022 11:30:52", "content": "

Just reached 0.59, 4fold single model, segformer backbone with UNet decoder.

\n

(Haven't done a HPA/HuBMAP prediction with it)

", "rawMarkdown": "Just reached 0.59, 4fold single model, segformer backbone with UNet decoder.\n\n(Haven't done a HPA/HuBMAP prediction with it)", "votes": null }, { "id": "1928713", "postDate": "09/06/2022 15:21:11", "content": "

Hi @kingjohnson Did you do that only by changing the threshold for each organs?
\nthanks in advance

", "rawMarkdown": "Hi @kingjohnson Did you do that only by changing the threshold for each organs? \nthanks in advance", "votes": null }, { "id": "1930110", "postDate": "09/07/2022 15:07:18", "content": "

Hi @jamesphoward , what is the training image size for your 4fold single model? 768 or 1024?

", "rawMarkdown": "Hi @jamesphoward , what is the training image size for your 4fold single model? 768 or 1024?", "votes": null }, { "id": "1930837", "postDate": "09/08/2022 08:41:58", "content": "

1024 pixels

", "rawMarkdown": "1024 pixels", "votes": null }, { "id": "1933165", "postDate": "09/10/2022 09:57:38", "content": "

I get 0.58 using 1 fold of single model, but my 4-fold ensemble score (4 of 5 folds) is even lower than 1 fold.😨

", "rawMarkdown": "I get 0.58 using 1 fold of single model, but my 4-fold ensemble score (4 of 5 folds) is even lower than 1 fold.😨", "votes": null }, { "id": "1933406", "postDate": "09/10/2022 13:39:50", "content": "

I try the method but it seems make no difference.

", "rawMarkdown": "I try the method but it seems make no difference.", "votes": null }, { "id": "1935260", "postDate": "09/12/2022 01:35:12", "content": "

Sorry for another question, what is your segformer backbone? (mit-b2? mit-b4?)

", "rawMarkdown": "Sorry for another question, what is your segformer backbone? (mit-b2? mit-b4?)", "votes": null }, { "id": "1935327", "postDate": "09/12/2022 03:22:50", "content": "

what is your organ threshold for each organ for Hubmap? thanks!

", "rawMarkdown": "what is your organ threshold for each organ for Hubmap? thanks!", "votes": null }, { "id": "1935743", "postDate": "09/12/2022 10:44:45", "content": "

I'm using Mit-B3

", "rawMarkdown": "I'm using Mit-B3", "votes": null }, { "id": "1936214", "postDate": "09/12/2022 15:14:26", "content": "

Thanks for your kind share! How did you conduct your 4-fold ensemble, average or weighted sum? During my ensemble, I found that simple average will reduce predicted segments and then degrade lb score.

", "rawMarkdown": "Thanks for your kind share! How did you conduct your 4-fold ensemble, average or weighted sum? During my ensemble, I found that simple average will reduce predicted segments and then degrade lb score.", "votes": null }, { "id": "1936221", "postDate": "09/12/2022 15:20:02", "content": "

I use the following, but this is very model specific and you will probably need different numbers:

\n

\"lung\": 0.05
\n\"kidney\": 0.45
\n\"largeintestine\": 0.3
\n\"prostate\": 0.25
\n\"spleen\": 0.25

\n

I use a simple average when ensembling.

", "rawMarkdown": "I use the following, but this is very model specific and you will probably need different numbers:\n\n\"lung\": 0.05\n\"kidney\": 0.45\n\"largeintestine\": 0.3\n\"prostate\": 0.25\n\"spleen\": 0.25\n\nI use a simple average when ensembling.", "votes": null }, { "id": "1936299", "postDate": "09/12/2022 16:37:25", "content": "

Thanks a lot!

", "rawMarkdown": "Thanks a lot!", "votes": null }, { "id": "1938709", "postDate": "09/14/2022 10:09:55", "content": "

My best single model (coat+daformer+768 + 5 folds ensemble) lb=0.60+ on HuBMAP.
\nBut we haven't broken 0.61 yet, I would like to know if anyone has lb=0.61+ on HuBMAP.

", "rawMarkdown": "My best single model (coat+daformer+768 + 5 folds ensemble) lb=0.60+ on HuBMAP.\nBut we haven't broken 0.61 yet, I would like to know if anyone has lb=0.61+ on HuBMAP.", "votes": null }, { "id": "1938721", "postDate": "09/14/2022 10:22:56", "content": "

Have you begun ensembling yet? If so, do you gain much performance? My best single model is 0.59 (4 folds). I can hit 0.60 with ensembling but really the benefit feels very small above 4 folds of my best model.

", "rawMarkdown": "Have you begun ensembling yet? If so, do you gain much performance? My best single model is 0.59 (4 folds). I can hit 0.60 with ensembling but really the benefit feels very small above 4 folds of my best model.", "votes": null }, { "id": "1938805", "postDate": "09/14/2022 11:21:47", "content": "

Our ensemble has improved but not much, it has not exceeded 0.61 on HuBMAP.

", "rawMarkdown": "Our ensemble has improved but not much, it has not exceeded 0.61 on HuBMAP.", "votes": null }, { "id": "1938812", "postDate": "09/14/2022 11:27:09", "content": "

It's too hard to get 0.61 on HuBMAP by single model, maybe I missed some important tricks or features. But I've got 0.61 by ensembling.

", "rawMarkdown": "It's too hard to get 0.61 on HuBMAP by single model, maybe I missed some important tricks or features. But I've got 0.61 by ensembling.", "votes": null }, { "id": "1939776", "postDate": "09/15/2022 01:57:40", "content": "

same here 0.58 🤓

", "rawMarkdown": "same here 0.58 🤓", "votes": null }, { "id": "1940202", "postDate": "09/15/2022 08:15:57", "content": "

May I ask what is your backbone model?(coat-lite small? coat lite medium?). And another question is that did you use Stain augmentation? Thanks!

", "rawMarkdown": "May I ask what is your backbone model?(coat-lite small? coat lite medium?). And another question is that did you use Stain augmentation? Thanks!", "votes": null }, { "id": "1940344", "postDate": "09/15/2022 09:52:56", "content": "

Gotta love this by \"Assistant Professor at Sejong University\"

", "rawMarkdown": "Gotta love this by \"Assistant Professor at Sejong University\"", "votes": null }, { "id": "1940346", "postDate": "09/15/2022 09:57:10", "content": "

@chris666 Hi,
\nBackbone with coat_lite_medium and use stain augmentation!

", "rawMarkdown": "chris666 Hi,\nBackbone with coat_lite_medium and use stain augmentation!", "votes": null }, { "id": "1940347", "postDate": "09/15/2022 09:58:30", "content": "

@rock139 Hi,
\nBut I've got 0.61 by ensembling.

\n

Glad to hear this, I will continue to improve my models.

", "rawMarkdown": "rock139 Hi, \n`But I've got 0.61 by ensembling.`\n\nGlad to hear this, I will continue to improve my models.", "votes": null } ], "comments": [ { "id": 1919549, "author_name": "chinartist", "author_url": "", "post_date": "08/30/2022 14:00:33", "content": "

Thanks for your sharing! Could you tell me how to know the my submission scores on HuBMAP?

", "votes": null, "replies": [ { "id": 1919568, "author_name": "gunesevitan", "author_url": "", "post_date": "08/30/2022 14:27:51", "content": "

Assign empty string to HPA samples.

", "votes": null, "replies": [] }, { "id": 1919754, "author_name": "soumya9977", "author_url": "", "post_date": "08/30/2022 16:47:18", "content": "

check this out https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333631#1835768

", "votes": null, "replies": [] } ] }, { "id": 1919688, "author_name": "benihime91", "author_url": "", "post_date": "08/30/2022 15:54:08", "content": "

same here 0.58 🤓

", "votes": null, "replies": [] }, { "id": 1920752, "author_name": "rock139", "author_url": "", "post_date": "08/31/2022 11:36:41", "content": "

My best submission scores 0.60 on HuBMAP，it's really hard to improve the performance on HuBMAP😂

", "votes": null, "replies": [ { "id": 1920899, "author_name": "sompark", "author_url": "", "post_date": "08/31/2022 13:11:36", "content": "

Amazing👍👍

", "votes": null, "replies": [] } ] }, { "id": 1921323, "author_name": "jamesphoward", "author_url": "", "post_date": "08/31/2022 18:02:42", "content": "

0.57 using 4 folds of deeplabV3-efficientnetB4.
\nIs your 0.58 and 0.59 ensembling, n-folds, or single model?

", "votes": null, "replies": [ { "id": 1921343, "author_name": "gunesevitan", "author_url": "", "post_date": "08/31/2022 18:42:23", "content": "

5 folds of single model.

", "votes": null, "replies": [] }, { "id": 1921793, "author_name": "chinartist", "author_url": "", "post_date": "09/01/2022 04:20:23", "content": "

Do you mean single model performs better?

", "votes": null, "replies": [] }, { "id": 1921903, "author_name": "gunesevitan", "author_url": "", "post_date": "09/01/2022 06:18:07", "content": "

I mean I'm using average of 5 folds of a model. Most of the time, that's what we refer to as a single model on Kaggle. I haven't started my ensemble yet. I'm still thinking about what kind of diversity should I bring with ensemble.

", "votes": null, "replies": [] } ] }, { "id": 1922018, "author_name": "befunny", "author_url": "", "post_date": "09/01/2022 08:04:13", "content": "

same as 0.59, 5fold of single model.

", "votes": null, "replies": [] }, { "id": 1922164, "author_name": "kingjohnson", "author_url": "", "post_date": "09/01/2022 10:06:02", "content": "

Why did my hubmap score stuck in 0.49? Is there anybody willing to help me find the problem?

", "votes": null, "replies": [ { "id": 1922182, "author_name": "jamesphoward", "author_url": "", "post_date": "09/01/2022 10:20:41", "content": "

Can you show your predictions on the single test image?

", "votes": null, "replies": [] }, { "id": 1922250, "author_name": "kingjohnson", "author_url": "", "post_date": "09/01/2022 11:27:38", "content": "

\"\"
\nDuring my local validation, in fold0, local cv reach 0.8269, which seems a good improvement. But lb only gets 0.70, and hubmap score is 0.49. It seems that my test results have many false positives.

", "votes": null, "replies": [] }, { "id": 1922376, "author_name": "rock139", "author_url": "", "post_date": "09/01/2022 12:50:34", "content": "

Maybe you can try to adjust the mask threshold in test images.

", "votes": null, "replies": [] }, { "id": 1924811, "author_name": "kingjohnson", "author_url": "", "post_date": "09/03/2022 12:34:34", "content": "

\"\"
\nMy new test result was presented above. Many false positives has been removed, and my hubmap score has achieved 0.52. Could anybody help me to further improve my score?

", "votes": null, "replies": [] }, { "id": 1928713, "author_name": "soumya9977", "author_url": "", "post_date": "09/06/2022 15:21:11", "content": "

Hi @kingjohnson Did you do that only by changing the threshold for each organs?
\nthanks in advance

", "votes": null, "replies": [] }, { "id": 1933406, "author_name": "kingjohnson", "author_url": "", "post_date": "09/10/2022 13:39:50", "content": "

I try the method but it seems make no difference.

", "votes": null, "replies": [] } ] }, { "id": 1922219, "author_name": "jiageng", "author_url": "", "post_date": "09/01/2022 10:55:58", "content": "

same as 0.59☝️

", "votes": null, "replies": [] }, { "id": 1928275, "author_name": "jamesphoward", "author_url": "", "post_date": "09/06/2022 11:30:52", "content": "

Just reached 0.59, 4fold single model, segformer backbone with UNet decoder.

\n

(Haven't done a HPA/HuBMAP prediction with it)

", "votes": null, "replies": [ { "id": 1935327, "author_name": "chris666", "author_url": "", "post_date": "09/12/2022 03:22:50", "content": "

what is your organ threshold for each organ for Hubmap? thanks!

", "votes": null, "replies": [] }, { "id": 1936214, "author_name": "kingjohnson", "author_url": "", "post_date": "09/12/2022 15:14:26", "content": "

Thanks for your kind share! How did you conduct your 4-fold ensemble, average or weighted sum? During my ensemble, I found that simple average will reduce predicted segments and then degrade lb score.

", "votes": null, "replies": [] }, { "id": 1936221, "author_name": "jamesphoward", "author_url": "", "post_date": "09/12/2022 15:20:02", "content": "

I use the following, but this is very model specific and you will probably need different numbers:

\n

\"lung\": 0.05
\n\"kidney\": 0.45
\n\"largeintestine\": 0.3
\n\"prostate\": 0.25
\n\"spleen\": 0.25

\n

I use a simple average when ensembling.

", "votes": null, "replies": [] } ] }, { "id": 1930110, "author_name": "chris666", "author_url": "", "post_date": "09/07/2022 15:07:18", "content": "

Hi @jamesphoward , what is the training image size for your 4fold single model? 768 or 1024?

", "votes": null, "replies": [ { "id": 1930837, "author_name": "jamesphoward", "author_url": "", "post_date": "09/08/2022 08:41:58", "content": "

1024 pixels

", "votes": null, "replies": [] }, { "id": 1935260, "author_name": "chris666", "author_url": "", "post_date": "09/12/2022 01:35:12", "content": "

Sorry for another question, what is your segformer backbone? (mit-b2? mit-b4?)

", "votes": null, "replies": [] }, { "id": 1935743, "author_name": "jamesphoward", "author_url": "", "post_date": "09/12/2022 10:44:45", "content": "

I'm using Mit-B3

", "votes": null, "replies": [] }, { "id": 1936299, "author_name": "chris666", "author_url": "", "post_date": "09/12/2022 16:37:25", "content": "

Thanks a lot!

", "votes": null, "replies": [] } ] }, { "id": 1933165, "author_name": "gufanmingmie", "author_url": "", "post_date": "09/10/2022 09:57:38", "content": "

I get 0.58 using 1 fold of single model, but my 4-fold ensemble score (4 of 5 folds) is even lower than 1 fold.😨

", "votes": null, "replies": [] }, { "id": 1938709, "author_name": "yingpengchen", "author_url": "", "post_date": "09/14/2022 10:09:55", "content": "

My best single model (coat+daformer+768 + 5 folds ensemble) lb=0.60+ on HuBMAP.
\nBut we haven't broken 0.61 yet, I would like to know if anyone has lb=0.61+ on HuBMAP.

", "votes": null, "replies": [ { "id": 1938721, "author_name": "jamesphoward", "author_url": "", "post_date": "09/14/2022 10:22:56", "content": "

Have you begun ensembling yet? If so, do you gain much performance? My best single model is 0.59 (4 folds). I can hit 0.60 with ensembling but really the benefit feels very small above 4 folds of my best model.

", "votes": null, "replies": [] }, { "id": 1938805, "author_name": "yingpengchen", "author_url": "", "post_date": "09/14/2022 11:21:47", "content": "

Our ensemble has improved but not much, it has not exceeded 0.61 on HuBMAP.

", "votes": null, "replies": [] }, { "id": 1938812, "author_name": "rock139", "author_url": "", "post_date": "09/14/2022 11:27:09", "content": "

It's too hard to get 0.61 on HuBMAP by single model, maybe I missed some important tricks or features. But I've got 0.61 by ensembling.

", "votes": null, "replies": [] }, { "id": 1940202, "author_name": "chris666", "author_url": "", "post_date": "09/15/2022 08:15:57", "content": "

May I ask what is your backbone model?(coat-lite small? coat lite medium?). And another question is that did you use Stain augmentation? Thanks!

", "votes": null, "replies": [] }, { "id": 1940346, "author_name": "yingpengchen", "author_url": "", "post_date": "09/15/2022 09:57:10", "content": "

@chris666 Hi,
\nBackbone with coat_lite_medium and use stain augmentation!

", "votes": null, "replies": [] }, { "id": 1940347, "author_name": "yingpengchen", "author_url": "", "post_date": "09/15/2022 09:58:30", "content": "

@rock139 Hi,
\nBut I've got 0.61 by ensembling.

\n

Glad to hear this, I will continue to improve my models.

", "votes": null, "replies": [] } ] }, { "id": 1939776, "author_name": "clyde1114", "author_url": "", "post_date": "09/15/2022 01:57:40", "content": "

same here 0.58 🤓

", "votes": null, "replies": [ { "id": 1940344, "author_name": "bakeryproducts", "author_url": "", "post_date": "09/15/2022 09:52:56", "content": "

Gotta love this by \"Assistant Professor at Sejong University\"

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1919500": "I decided to submit only HuBMAP after reaching 0.8 since I want to use my last submissions efficiently. My best submission scores 0.58 on HuBMAP only part of the test set.\n\n**Edit**: I just reached 0.59. I guess 0.62 is the gold zone.\n**Edit 2**: I overestimated the gold zone. I'm still struggling to reach 0.60.", "1919549": "Thanks for your sharing! Could you tell me how to know the my submission scores on HuBMAP?", "1919568": "Assign empty string to HPA samples.", "1919688": "same here 0.58 🤓", "1919754": "check this out https://www.kaggle.com/competitions/hubmap-organ-segmentation/discussion/333631#1835768", "1920752": "My best submission scores 0.60 on HuBMAP，it's really hard to improve the performance on HuBMAP😂", "1920899": "Amazing👍👍", "1921323": "0.57 using 4 folds of deeplabV3-efficientnetB4.\nIs your 0.58 and 0.59 ensembling, n-folds, or single model?", "1921343": "5 folds of single model.", "1921793": "Do you mean single model performs better?", "1921903": "I mean I'm using average of 5 folds of a model. Most of the time, that's what we refer to as a single model on Kaggle. I haven't started my ensemble yet. I'm still thinking about what kind of diversity should I bring with ensemble.", "1922018": "same as 0.59, 5fold of single model.", "1922164": "Why did my hubmap score stuck in 0.49? Is there anybody willing to help me find the problem?", "1922182": "Can you show your predictions on the single test image?", "1922219": "same as 0.59☝️", "1922250": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6149314%2F7e296638ef80efff944d0b797aeea0a6%2F10078_050.jpg?generation=1662031531847507&alt=media)\nDuring my local validation, in fold0, local cv reach 0.8269, which seems a good improvement. But lb only gets 0.70, and hubmap score is 0.49. It seems that my test results have many false positives.", "1922376": "Maybe you can try to adjust the mask threshold in test images.", "1924811": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F6149314%2F786878a0b0ee82545e519449d4cecfdb%2F10078_092_b5_040.jpg?generation=1662208416757805&alt=media)\nMy new test result was presented above. Many false positives has been removed, and my hubmap score has achieved 0.52. Could anybody help me to further improve my score?", "1928275": "Just reached 0.59, 4fold single model, segformer backbone with UNet decoder.\n\n(Haven't done a HPA/HuBMAP prediction with it)", "1928713": "Hi @kingjohnson Did you do that only by changing the threshold for each organs? \nthanks in advance", "1930110": "Hi @jamesphoward , what is the training image size for your 4fold single model? 768 or 1024?", "1930837": "1024 pixels", "1933165": "I get 0.58 using 1 fold of single model, but my 4-fold ensemble score (4 of 5 folds) is even lower than 1 fold.😨", "1933406": "I try the method but it seems make no difference.", "1935260": "Sorry for another question, what is your segformer backbone? (mit-b2? mit-b4?)", "1935327": "what is your organ threshold for each organ for Hubmap? thanks!", "1935743": "I'm using Mit-B3", "1936214": "Thanks for your kind share! How did you conduct your 4-fold ensemble, average or weighted sum? During my ensemble, I found that simple average will reduce predicted segments and then degrade lb score.", "1936221": "I use the following, but this is very model specific and you will probably need different numbers:\n\n\"lung\": 0.05\n\"kidney\": 0.45\n\"largeintestine\": 0.3\n\"prostate\": 0.25\n\"spleen\": 0.25\n\nI use a simple average when ensembling.", "1936299": "Thanks a lot!", "1938709": "My best single model (coat+daformer+768 + 5 folds ensemble) lb=0.60+ on HuBMAP.\nBut we haven't broken 0.61 yet, I would like to know if anyone has lb=0.61+ on HuBMAP.", "1938721": "Have you begun ensembling yet? If so, do you gain much performance? My best single model is 0.59 (4 folds). I can hit 0.60 with ensembling but really the benefit feels very small above 4 folds of my best model.", "1938805": "Our ensemble has improved but not much, it has not exceeded 0.61 on HuBMAP.", "1938812": "It's too hard to get 0.61 on HuBMAP by single model, maybe I missed some important tricks or features. But I've got 0.61 by ensembling.", "1939776": "same here 0.58 🤓", "1940202": "May I ask what is your backbone model?(coat-lite small? coat lite medium?). And another question is that did you use Stain augmentation? Thanks!", "1940344": "Gotta love this by \"Assistant Professor at Sejong University\"", "1940346": "chris666 Hi,\nBackbone with coat_lite_medium and use stain augmentation!", "1940347": "rock139 Hi, \n`But I've got 0.61 by ensembling.`\n\nGlad to hear this, I will continue to improve my models." }, "source": "meta" }