{ "id": 354744, "title": "25th Simple Solution", "url": "/competitions/hubmap-organ-segmentation/discussion/354744", "author_name": "kaggler", "post_date": "2022-09-23T15:16:13.317000", "votes": 20, "comment_count": 6, "views": 0, "content": "

I would like to thank the organizers for hosting the great competition.
\nalso, I would like to express my gratitude to our teammate @hwigeon and new teammate @methyl for their dedication to the competition.
\nI also want to express my greatest gratitude to the frog @hengck23. Without his comments through the open discusstion, I would have given up this competition.

\n

Augmentations

\n

I can't cross even 0.6 without insane augmentations.
\nAfter I applied very strong augmentations with a lot of epochs, I could stably score over 0.75+

\n

I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning
\nafter these augmentations, we can score over 0.75 with most of our models

\n

Models

\n

Kaggler : ( Patch + Finetune ) Efficientnet7+Deeplabv3plus, Efficientnet6+Deeplabv3plus, Efficientnetv2l+Deeplabv3Plus, Coat-small + Coat-small-pl [768,1536, 1024,1536,1536]
\nHwigeon : (Only Patch) swin transformer_Unet + PvT + Coat
\nMethyl : (Only whole image) PVT + Coat [1024,1024,1024,1024]

\n

Validation

\n

We failed validation. When teaming up, though I know validation is very important, but with time being constrained or as the performances of some models are not totally reproduced, we can't validate our cv properly. We just depended on hubmap LB.

\n

Threshold

\n

organ_threshold = {
\n 'Hubmap': {
\n 'kidney' : 0.40,
\n 'prostate' : 0.40,
\n 'largeintestine': 0.40,
\n 'spleen' : 0.40,
\n 'lung' : 0.10,
\n }}

\n

We failed to get the gold medal but There is no regret since I tried my best
\nThank you for reading this!

", "messages": [ { "id": 1952333, "postDate": "2022-09-23T15:16:13.317Z", "content": "

I would like to thank the organizers for hosting the great competition.
\nalso, I would like to express my gratitude to our teammate @hwigeon and new teammate @methyl for their dedication to the competition.
\nI also want to express my greatest gratitude to the frog @hengck23. Without his comments through the open discusstion, I would have given up this competition.

\n

Augmentations

\n

I can't cross even 0.6 without insane augmentations.
\nAfter I applied very strong augmentations with a lot of epochs, I could stably score over 0.75+

\n

I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning
\nafter these augmentations, we can score over 0.75 with most of our models

\n

Models

\n

Kaggler : ( Patch + Finetune ) Efficientnet7+Deeplabv3plus, Efficientnet6+Deeplabv3plus, Efficientnetv2l+Deeplabv3Plus, Coat-small + Coat-small-pl [768,1536, 1024,1536,1536]
\nHwigeon : (Only Patch) swin transformer_Unet + PvT + Coat
\nMethyl : (Only whole image) PVT + Coat [1024,1024,1024,1024]

\n

Validation

\n

We failed validation. When teaming up, though I know validation is very important, but with time being constrained or as the performances of some models are not totally reproduced, we can't validate our cv properly. We just depended on hubmap LB.

\n

Threshold

\n

organ_threshold = {
\n 'Hubmap': {
\n 'kidney' : 0.40,
\n 'prostate' : 0.40,
\n 'largeintestine': 0.40,
\n 'spleen' : 0.40,
\n 'lung' : 0.10,
\n }}

\n

We failed to get the gold medal but There is no regret since I tried my best
\nThank you for reading this!

", "rawMarkdown": "I would like to thank the organizers for hosting the great competition.\nalso, I would like to express my gratitude to our teammate @hwigeon and new teammate @methyl for their dedication to the competition.\nI also want to express my greatest gratitude to the frog @hengck23. Without his comments through the open discusstion, I would have given up this competition. \n \n# Augmentations\nI can't cross even 0.6 without insane augmentations. \nAfter I applied very strong augmentations with a lot of epochs, I could stably score over 0.75+ \n\nI applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning\nafter these augmentations, we can score over 0.75 with most of our models\n\n# Models \nKaggler : ( Patch + Finetune ) Efficientnet7+Deeplabv3plus, Efficientnet6+Deeplabv3plus, Efficientnetv2l+Deeplabv3Plus, Coat-small + Coat-small-pl [768,1536, 1024,1536,1536]\nHwigeon : (Only Patch) swin transformer_Unet + PvT + Coat\nMethyl : (Only whole image) PVT + Coat [1024,1024,1024,1024]\n \n\n# Validation \nWe failed validation. When teaming up, though I know validation is very important, but with time being constrained or as the performances of some models are not totally reproduced, we can't validate our cv properly. We just depended on hubmap LB. \n\n# Threshold\norgan_threshold = {\n 'Hubmap': {\n 'kidney' : 0.40,\n 'prostate' : 0.40,\n 'largeintestine': 0.40,\n 'spleen' : 0.40,\n 'lung' : 0.10,\n }}\n \nWe failed to get the gold medal but There is no regret since I tried my best \nThank you for reading this!", "votes": 20 }, { "id": 1952628, "postDate": "2022-09-23T19:43:36.133Z", "content": "

try to make another submission with threshold lower, e.g.
\norgan_threshold = {
\n'Hubmap': {
\n'kidney' : 0.30,
\n'prostate' : 0.30,
\n'largeintestine': 0.30,
\n'spleen' : 0.30,
\n'lung' : 0.10,
\n}}

\n

as you ensemble, it is possible that the ensemble has better ROC/precise-recall curve and the thtreshold can be lowered

", "rawMarkdown": "try to make another submission with threshold lower, e.g. \norgan_threshold = {\n'Hubmap': {\n'kidney' : 0.30,\n'prostate' : 0.30,\n'largeintestine': 0.30,\n'spleen' : 0.30,\n'lung' : 0.10,\n}}\n\n\nas you ensemble, it is possible that the ensemble has better ROC/precise-recall curve and the thtreshold can be lowered ", "votes": 2, "replies": [ { "id": 1952801, "postDate": "2022-09-23T23:59:50.800Z", "content": "

Why do better ROC / precise-recall cure mean that threshold can be lowered, can you further explain about it?Thanks a lot for your kind share.

", "rawMarkdown": "Why do better ROC / precise-recall cure mean that threshold can be lowered, can you further explain about it?Thanks a lot for your kind share.", "votes": 1 }, { "id": 1953362, "postDate": "2022-09-24T12:24:02.990Z", "content": "

@hengck23 which one is better, increasing thresholds or lowering thresholds?
\nisn't it the same situation when more and more models are being ensembled

", "rawMarkdown": "@hengck23 which one is better, increasing thresholds or lowering thresholds? \nisn't it the same situation when more and more models are being ensembled" }, { "id": 1953678, "postDate": "2022-09-24T16:49:04.420Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1954248, "postDate": "2022-09-25T05:57:29.560Z", "rawMarkdown": "", "votes": -1, "isDeleted": true } ] }, { "id": 1972235, "postDate": "2022-10-05T04:13:51.133Z", "content": "

Can you share the more detail of your insane augmentation

\n
\n

I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning
\n after these augmentations, we can score over 0.75 with most of our models

\n
\n

I tried strong augmentation but it didn't work.
\nI'm curious about your augmentation parameter.

", "rawMarkdown": "Can you share the more detail of your insane augmentation\n\n> I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning\nafter these augmentations, we can score over 0.75 with most of our models\n\nI tried strong augmentation but it didn't work.\nI'm curious about your augmentation parameter." } ], "comments": [ { "id": 1952628, "author_name": "hengck23", "author_url": "", "post_date": "2022-09-23T19:43:36.133000", "content": "

try to make another submission with threshold lower, e.g.
\norgan_threshold = {
\n'Hubmap': {
\n'kidney' : 0.30,
\n'prostate' : 0.30,
\n'largeintestine': 0.30,
\n'spleen' : 0.30,
\n'lung' : 0.10,
\n}}

\n

as you ensemble, it is possible that the ensemble has better ROC/precise-recall curve and the thtreshold can be lowered

", "votes": 2, "replies": [ { "id": 1952801, "author_name": "kingjohnson", "author_url": "", "post_date": "2022-09-23T23:59:50.800000", "content": "

Why do better ROC / precise-recall cure mean that threshold can be lowered, can you further explain about it?Thanks a lot for your kind share.

", "votes": 1, "replies": [] }, { "id": 1953362, "author_name": "kaggler", "author_url": "", "post_date": "2022-09-24T12:24:02.990000", "content": "

@hengck23 which one is better, increasing thresholds or lowering thresholds?
\nisn't it the same situation when more and more models are being ensembled

", "votes": 0, "replies": [] }, { "id": 1953678, "author_name": "", "author_url": "", "post_date": "2022-09-24T16:49:04.420000", "content": "", "votes": 1, "replies": [] }, { "id": 1954248, "author_name": "", "author_url": "", "post_date": "2022-09-25T05:57:29.560000", "content": "", "votes": -1, "replies": [] } ] }, { "id": 1972235, "author_name": "yoshoo", "author_url": "", "post_date": "2022-10-05T04:13:51.133000", "content": "

Can you share the more detail of your insane augmentation

\n
\n

I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning
\n after these augmentations, we can score over 0.75 with most of our models

\n
\n

I tried strong augmentation but it didn't work.
\nI'm curious about your augmentation parameter.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1952333": "I would like to thank the organizers for hosting the great competition.\nalso, I would like to express my gratitude to our teammate @hwigeon and new teammate @methyl for their dedication to the competition.\nI also want to express my greatest gratitude to the frog @hengck23. Without his comments through the open discusstion, I would have given up this competition. \n \n# Augmentations\nI can't cross even 0.6 without insane augmentations. \nAfter I applied very strong augmentations with a lot of epochs, I could stably score over 0.75+ \n\nI applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning\nafter these augmentations, we can score over 0.75 with most of our models\n\n# Models \nKaggler : ( Patch + Finetune ) Efficientnet7+Deeplabv3plus, Efficientnet6+Deeplabv3plus, Efficientnetv2l+Deeplabv3Plus, Coat-small + Coat-small-pl [768,1536, 1024,1536,1536]\nHwigeon : (Only Patch) swin transformer_Unet + PvT + Coat\nMethyl : (Only whole image) PVT + Coat [1024,1024,1024,1024]\n \n\n# Validation \nWe failed validation. When teaming up, though I know validation is very important, but with time being constrained or as the performances of some models are not totally reproduced, we can't validate our cv properly. We just depended on hubmap LB. \n\n# Threshold\norgan_threshold = {\n 'Hubmap': {\n 'kidney' : 0.40,\n 'prostate' : 0.40,\n 'largeintestine': 0.40,\n 'spleen' : 0.40,\n 'lung' : 0.10,\n }}\n \nWe failed to get the gold medal but There is no regret since I tried my best \nThank you for reading this!", "1952628": "try to make another submission with threshold lower, e.g. \norgan_threshold = {\n'Hubmap': {\n'kidney' : 0.30,\n'prostate' : 0.30,\n'largeintestine': 0.30,\n'spleen' : 0.30,\n'lung' : 0.10,\n}}\n\n\nas you ensemble, it is possible that the ensemble has better ROC/precise-recall curve and the thtreshold can be lowered ", "1972235": "Can you share the more detail of your insane augmentation\n\n> I applied HSV, BRIGHTNESS, RANDOMCROPANDRESIZE, STAIN NORMALIZATION, AUGRESSIVE SIZE CHANGE OF PROSTATE with 500 epoch patch training + 400 epoch whole region finetuning\nafter these augmentations, we can score over 0.75 with most of our models\n\nI tried strong augmentation but it didn't work.\nI'm curious about your augmentation parameter." } }