{ "id": 226576, "title": "Our CV 0.976057 Ensemble Wins!", "url": "/competitions/ranzcr-clip-catheter-line-classification/discussion/226576", "author_name": "Qishen Ha", "post_date": "2021-03-17T01:58:55.859000", "votes": 80, "comment_count": 41, "views": 0, "content": "

Congratulations to all winners and thanks to the kaggle, the host and my wonderful teammates! @boliu0 @garybios

\n

Our winning submission is based on 5 segmentation models (1024x1024 ~ 1536x1536) and 67 classification models (384x384 ~ 512x512), which took 8.7h to run.

\n

Among our 22 submissions, this one has the 3rd highest pvt score but with the highest local cv —— 0.976057

\n

The cv strategy we use is Stratified Group K Fold.

\n

Brief summary of our solution has been released by @boliu0 :
\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633

\n

Furthermore, I've released our minimal pipeline here: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/227128 which is able to reach pub/pvt LB 0.972+ by training only on a single 16GB GPU.

", "messages": [ { "id": 1241316, "postDate": "2021-03-17T01:58:55.860Z", "content": "

Congratulations to all winners and thanks to the kaggle, the host and my wonderful teammates! @boliu0 @garybios

\n

Our winning submission is based on 5 segmentation models (1024x1024 ~ 1536x1536) and 67 classification models (384x384 ~ 512x512), which took 8.7h to run.

\n

Among our 22 submissions, this one has the 3rd highest pvt score but with the highest local cv —— 0.976057

\n

The cv strategy we use is Stratified Group K Fold.

\n

Brief summary of our solution has been released by @boliu0 :
\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633

\n

Furthermore, I've released our minimal pipeline here: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/227128 which is able to reach pub/pvt LB 0.972+ by training only on a single 16GB GPU.

", "rawMarkdown": "Congratulations to all winners and thanks to the kaggle, the host and my wonderful teammates! @boliu0 @garybios\n\nOur winning submission is based on 5 segmentation models (1024x1024 ~ 1536x1536) and 67 classification models (384x384 ~ 512x512), which took 8.7h to run.\n\nAmong our 22 submissions, this one has the 3rd highest pvt score but with the highest local cv —— **0.976057**\n\nThe cv strategy we use is Stratified Group K Fold.\n\nBrief summary of our solution has been released by @boliu0 :\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633\n\nFurthermore, I've released our minimal pipeline here: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/227128 which is able to reach **pub/pvt LB 0.972+** by training only on a single 16GB GPU.", "votes": 80 }, { "id": 1241470, "postDate": "2021-03-17T04:05:08.193Z", "content": "

Congrats )) great work. Well I think there is no need anymore for classification competitions they can put you always at the top and end the competition XD XD

", "rawMarkdown": "Congrats )) great work. Well I think there is no need anymore for classification competitions they can put you always at the top and end the competition XD XD", "votes": 7 }, { "id": 1241459, "postDate": "2021-03-17T03:49:50.670Z", "content": "

Amazing, How can you train 67 classification model? By automatic searching different hyperparameter or purely by hand adjust?

", "rawMarkdown": "Amazing, How can you train 67 classification model? By automatic searching different hyperparameter or purely by hand adjust?", "votes": 3, "replies": [ { "id": 1241511, "postDate": "2021-03-17T04:36:55Z", "content": "

We designed models with good diversity by hand and trained them, and then ensemble oofs to finally select some of them.

", "rawMarkdown": "We designed models with good diversity by hand and trained them, and then ensemble oofs to finally select some of them.", "votes": 5 }, { "id": 1241577, "postDate": "2021-03-17T05:33:28.290Z", "content": "

it's actually 31 models, but we selected 67 folds from these 31 models (i.e. 3 folds from some models, 1 fold from some other models) for the inference kernel

", "rawMarkdown": "it's actually 31 models, but we selected 67 folds from these 31 models (i.e. 3 folds from some models, 1 fold from some other models) for the inference kernel", "votes": 4 } ] }, { "id": 1241409, "postDate": "2021-03-17T03:07:00.173Z", "content": "

congratulation, it's too crazy, 67 classification models and so big input segment model!!!

", "rawMarkdown": "congratulation, it's too crazy, 67 classification models and so big input segment model!!!", "votes": 3, "replies": [ { "id": 1241589, "postDate": "2021-03-17T05:37:49.857Z", "content": "

Thanks. We intentionally trained many small classification models, to balance good diversity and limited inference time. With segmentation output having very high quality, small and large classification models' performance are not very big.

", "rawMarkdown": "Thanks. We intentionally trained many small classification models, to balance good diversity and limited inference time. With segmentation output having very high quality, small and large classification models' performance are not very big.", "votes": 2 } ] }, { "id": 1241339, "postDate": "2021-03-17T02:14:38.130Z", "content": "

congratulation!

", "rawMarkdown": "congratulation!", "votes": 3 }, { "id": 1241333, "postDate": "2021-03-17T02:09:09.917Z", "content": "

Congratulations!🎉🎉🎉
\nIt's very impressive to win a competition with only 22 submissions. Even the LB is stable.

", "rawMarkdown": "Congratulations!🎉🎉🎉 \nIt's very impressive to win a competition with only 22 submissions. Even the LB is stable.", "votes": 3 }, { "id": 1241324, "postDate": "2021-03-17T02:03:44.187Z", "content": "
\n

5 segmentation models (1024x1024 ~ 1536~1536) and 67 classification models (384x384 ~ 512x512)

\n
\n

my god. but it's really creative you guys use big size for segmentation and only small size for classification. congrats!

", "rawMarkdown": "> 5 segmentation models (1024x1024 ~ 1536~1536) and 67 classification models (384x384 ~ 512x512)\n\nmy god. but it's really creative you guys use big size for segmentation and only small size for classification. congrats!", "votes": 3, "replies": [ { "id": 1241342, "postDate": "2021-03-17T02:17:01.103Z", "content": "

can we confirm that although there is an advantage for increasing image size classification, but there is a limit.

\n

beyond that, there is no more improvement for using large size image, even if you increase the parameters of the classifier.

\n

then the only to improve is the switch to segmentation for large image?

\n
\n

and i would also wonder what is the effect of using small image for segmentation (or feature from segmentation pretraining)?

", "rawMarkdown": "can we confirm that although there is an advantage for increasing image size classification, but there is a limit. \n\nbeyond that, there is no more improvement for using large size image, even if you increase the parameters of the classifier.\n\nthen the only to improve is the switch to segmentation for large image?\n\n---\nand i would also wonder what is the effect of using small image for segmentation (or feature from segmentation pretraining)?", "votes": 2 }, { "id": 1241345, "postDate": "2021-03-17T02:19:26.650Z", "content": "

yes, we found that the quality of the mask determines the performance of the classification, so we used big size for training the segmentation model to obtain a higher quality mask.

", "rawMarkdown": "yes, we found that the quality of the mask determines the performance of the classification, so we used big size for training the segmentation model to obtain a higher quality mask.", "votes": 7 } ] }, { "id": 1241338, "postDate": "2021-03-17T02:13:44.867Z", "content": "

congratulation!

", "rawMarkdown": "congratulation!", "votes": 4 }, { "id": 1244823, "postDate": "2021-03-19T08:45:45.417Z", "content": "

Congratulations

", "rawMarkdown": "Congratulations", "votes": 1 }, { "id": 1244348, "postDate": "2021-03-19T00:01:26.367Z", "content": "

Congrats ! Looking forward to your solution

", "rawMarkdown": "Congrats ! Looking forward to your solution ", "votes": 1 }, { "id": 1244318, "postDate": "2021-03-18T22:40:15.683Z", "content": "

congratulation！！！

", "rawMarkdown": "congratulation！！！", "votes": 1 }, { "id": 1242410, "postDate": "2021-03-17T15:41:30.037Z", "content": "

That's a lot of models. 😁
\nCongrats. 🤘

", "rawMarkdown": "That's a lot of models. 😁 \nCongrats. 🤘 ", "votes": 1 }, { "id": 1242142, "postDate": "2021-03-17T12:34:16.163Z", "content": "

Congratulation! It's amazing how few commits there are!

", "rawMarkdown": "Congratulation! It's amazing how few commits there are!", "votes": 1 }, { "id": 1241676, "postDate": "2021-03-17T06:40:05.430Z", "content": "

congratulation!

", "rawMarkdown": "congratulation!", "votes": 1 }, { "id": 1241665, "postDate": "2021-03-17T06:30:36.883Z", "content": "

Congrats on 1st place and great team effort!
\n@haqishen @boliu0 @garybios

\n

I'm looking forward to see your team solution
\nAnd thank you for always releasing minimal pipeline.

\n

p.s. I still remember the Melanoma competition. :)

", "rawMarkdown": "Congrats on 1st place and great team effort!\n@haqishen @boliu0 @garybios \n\nI'm looking forward to see your team solution\nAnd thank you for always releasing minimal pipeline.\n\np.s. I still remember the Melanoma competition. :)", "votes": 1 }, { "id": 1241476, "postDate": "2021-03-17T04:12:03.327Z", "content": "

Congratulations on 1st place!! CV0.976 is very impressive.

", "rawMarkdown": "Congratulations on 1st place!! CV0.976 is very impressive.", "votes": 1 }, { "id": 1241431, "postDate": "2021-03-17T03:23:41.833Z", "content": "

Congratulations. Fantastic effort

", "rawMarkdown": "Congratulations. Fantastic effort", "votes": 1 }, { "id": 1241426, "postDate": "2021-03-17T03:20:23.103Z", "content": "

Congratulations on the win @haqishen

", "rawMarkdown": "Congratulations on the win @haqishen ", "votes": 1 }, { "id": 1241360, "postDate": "2021-03-17T02:33:35.210Z", "content": "

Strong Ensemble and solid CV. Well done congrats @haqishen @garybios and @boliu0. Looking forward to reading detail solution

", "rawMarkdown": "Strong Ensemble and solid CV. Well done congrats @haqishen @garybios and @boliu0. Looking forward to reading detail solution", "votes": 1 }, { "id": 1241344, "postDate": "2021-03-17T02:19:16.110Z", "content": "

Too strong!!! Congrats!!!!!

", "rawMarkdown": "Too strong!!! Congrats!!!!!", "votes": 1 }, { "id": 1241340, "postDate": "2021-03-17T02:15:00.650Z", "content": "

We had a CV of 0.9767 but only scored 0.971 at LB. i think we had a data leak

", "rawMarkdown": "We had a CV of 0.9767 but only scored 0.971 at LB. i think we had a data leak", "votes": 1, "replies": [ { "id": 1241348, "postDate": "2021-03-17T02:21:12.187Z", "content": "

You used pretrained weights from the forum right?

", "rawMarkdown": "You used pretrained weights from the forum right?", "votes": 2 }, { "id": 1241353, "postDate": "2021-03-17T02:27:45.663Z", "content": "

yes . but scored 0.972 at private which we combined with our best ensemble to make our select submission

", "rawMarkdown": "yes . but scored 0.972 at private which we combined with our best ensemble to make our select submission" }, { "id": 1241365, "postDate": "2021-03-17T02:35:27.693Z", "content": "

When use such pretrained weights you should be cafeful on what data have been used to train it.

\n

That says, If sample A is trained in the pretrained weight, data leak will occur if sample A is validated in your finetune process.

", "rawMarkdown": "When use such pretrained weights you should be cafeful on what data have been used to train it.\n\nThat says, If sample A is trained in the pretrained weight, data leak will occur if sample A is validated in your finetune process.", "votes": 4 }, { "id": 1241377, "postDate": "2021-03-17T02:38:52.723Z", "content": "

When i recheck our submission . 0.9767 was a CV of single model without using any of the pretrained weight of the forum.

", "rawMarkdown": "When i recheck our submission . 0.9767 was a CV of single model without using any of the pretrained weight of the forum.", "votes": 1 }, { "id": 1241391, "postDate": "2021-03-17T02:50:02.240Z", "rawMarkdown": "", "isDeleted": true }, { "id": 1241390, "postDate": "2021-03-17T02:50:02.240Z", "content": "

Hope you can find out what caused this gap… 😂

", "rawMarkdown": "Hope you can find out what caused this gap... 😂", "votes": 4 }, { "id": 1241439, "postDate": "2021-03-17T03:30:59.420Z", "content": "

i also think there are leaks on your cv. my local cv scores 0.974 and private lb scores 0.975

", "rawMarkdown": "i also think there are leaks on your cv. my local cv scores 0.974 and private lb scores 0.975", "votes": 1 } ] }, { "id": 1246766, "postDate": "2021-03-21T04:18:37.680Z", "content": "

Huge Congratulations!👍

", "rawMarkdown": "Huge Congratulations!👍", "votes": 2 }, { "id": 1241777, "postDate": "2021-03-17T07:38:03.333Z", "content": "

Congratz ! An impressive result once again. I'll definitely study your solution carefully :)

", "rawMarkdown": "Congratz ! An impressive result once again. I'll definitely study your solution carefully :)", "votes": 2 }, { "id": 1241375, "postDate": "2021-03-17T02:38:22.340Z", "content": "
\n

5 segmentation models (1024x1024 ~ 1536x1536)

\n
\n

What kind of hardware do you use to work with this size?
\nCongratulations on the win !

", "rawMarkdown": "> 5 segmentation models (1024x1024 ~ 1536x1536)\n\nWhat kind of hardware do you use to work with this size?\nCongratulations on the win ! ", "votes": 2, "replies": [ { "id": 1241397, "postDate": "2021-03-17T02:58:12.280Z", "content": "

Thanks!

\n

A month ago, I trained Unet-B5 1024x1024 on a single NVIDIA Quadro RTX6000 with 24GB and combine it with a B3 512 classification model to get CV 0.969 (5 fold seg + 5 fold cls ensemble gets pub and pvt LB 0.972)

\n

After that we didn't submit for a long time.

\n

You didn't see us on the LB before because we only predicted about 90% of the public test data at that time.

\n

I'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.

", "rawMarkdown": "Thanks!\n\nA month ago, I trained Unet-B5 1024x1024 on a single **NVIDIA Quadro RTX6000 with 24GB** and combine it with a B3 512 classification model to get CV 0.969 (5 fold seg + 5 fold cls ensemble gets pub and pvt LB 0.972)\n\nAfter that we didn't submit for a long time.\n\nYou didn't see us on the LB before because we only predicted about 90% of the public test data at that time.\n\nI'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.", "votes": 8 }, { "id": 1242639, "postDate": "2021-03-17T18:15:47.510Z", "content": "

That would be very interesting, thanks in advance:

\n
\n

I'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.

\n
", "rawMarkdown": "That would be very interesting, thanks in advance: \n\n> I'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.", "votes": 3 } ] }, { "id": 1241374, "postDate": "2021-03-17T02:37:57.783Z", "content": "

congratulation!

", "rawMarkdown": "congratulation!", "votes": 2 }, { "id": 1241321, "postDate": "2021-03-17T02:01:46.673Z", "content": "

Congratz !!!!

\n
\n

which took 8.7h to run.

\n
\n

just too close to timeup

", "rawMarkdown": "Congratz !!!! \n>which took 8.7h to run.\n\njust too close to timeup", "votes": 2 }, { "id": 1241337, "postDate": "2021-03-17T02:12:18.373Z", "content": "

@haqishen Congratulations on 1 st place Finish with too few submissions

", "rawMarkdown": "@haqishen Congratulations on 1 st place Finish with too few submissions", "votes": 1 }, { "id": 1248949, "postDate": "2021-03-23T00:41:49.170Z", "content": "

Congratulations!🎉🎉

", "rawMarkdown": "Congratulations!🎉🎉" } ], "comments": [ { "id": 1241470, "author_name": "ammarali32", "author_url": "", "post_date": "2021-03-17T04:05:08.193000", "content": "

Congrats )) great work. Well I think there is no need anymore for classification competitions they can put you always at the top and end the competition XD XD

", "votes": 7, "replies": [] }, { "id": 1241459, "author_name": "sheep", "author_url": "", "post_date": "2021-03-17T03:49:50.670000", "content": "

Amazing, How can you train 67 classification model? By automatic searching different hyperparameter or purely by hand adjust?

", "votes": 3, "replies": [ { "id": 1241511, "author_name": "Qishen Ha", "author_url": "", "post_date": "2021-03-17T04:36:55", "content": "

We designed models with good diversity by hand and trained them, and then ensemble oofs to finally select some of them.

", "votes": 5, "replies": [] }, { "id": 1241577, "author_name": "Bo", "author_url": "", "post_date": "2021-03-17T05:33:28.290000", "content": "

it's actually 31 models, but we selected 67 folds from these 31 models (i.e. 3 folds from some models, 1 fold from some other models) for the inference kernel

", "votes": 4, "replies": [] } ] }, { "id": 1241409, "author_name": "cswwp", "author_url": "", "post_date": "2021-03-17T03:07:00.173000", "content": "

congratulation, it's too crazy, 67 classification models and so big input segment model!!!

", "votes": 3, "replies": [ { "id": 1241589, "author_name": "Bo", "author_url": "", "post_date": "2021-03-17T05:37:49.857000", "content": "

Thanks. We intentionally trained many small classification models, to balance good diversity and limited inference time. With segmentation output having very high quality, small and large classification models' performance are not very big.

", "votes": 2, "replies": [] } ] }, { "id": 1241339, "author_name": "sin", "author_url": "", "post_date": "2021-03-17T02:14:38.130000", "content": "

congratulation!

", "votes": 3, "replies": [] }, { "id": 1241333, "author_name": "toxu", "author_url": "", "post_date": "2021-03-17T02:09:09.917000", "content": "

Congratulations!🎉🎉🎉
\nIt's very impressive to win a competition with only 22 submissions. Even the LB is stable.

", "votes": 3, "replies": [] }, { "id": 1241324, "author_name": "DatNT", "author_url": "", "post_date": "2021-03-17T02:03:44.187000", "content": "
\n

5 segmentation models (1024x1024 ~ 1536~1536) and 67 classification models (384x384 ~ 512x512)

\n
\n

my god. but it's really creative you guys use big size for segmentation and only small size for classification. congrats!

", "votes": 3, "replies": [ { "id": 1241342, "author_name": "hengck23", "author_url": "", "post_date": "2021-03-17T02:17:01.103000", "content": "

can we confirm that although there is an advantage for increasing image size classification, but there is a limit.

\n

beyond that, there is no more improvement for using large size image, even if you increase the parameters of the classifier.

\n

then the only to improve is the switch to segmentation for large image?

\n
\n

and i would also wonder what is the effect of using small image for segmentation (or feature from segmentation pretraining)?

", "votes": 2, "replies": [] }, { "id": 1241345, "author_name": "Gary", "author_url": "", "post_date": "2021-03-17T02:19:26.650000", "content": "

yes, we found that the quality of the mask determines the performance of the classification, so we used big size for training the segmentation model to obtain a higher quality mask.

", "votes": 7, "replies": [] } ] }, { "id": 1241338, "author_name": "hengck23", "author_url": "", "post_date": "2021-03-17T02:13:44.867000", "content": "

congratulation!

", "votes": 4, "replies": [] }, { "id": 1244823, "author_name": "Archana benur", "author_url": "", "post_date": "2021-03-19T08:45:45.417000", "content": "

Congratulations

", "votes": 1, "replies": [] }, { "id": 1244348, "author_name": "Sidney Ng", "author_url": "", "post_date": "2021-03-19T00:01:26.367000", "content": "

Congrats ! Looking forward to your solution

", "votes": 1, "replies": [] }, { "id": 1244318, "author_name": "gaki ina", "author_url": "", "post_date": "2021-03-18T22:40:15.683000", "content": "

congratulation！！！

", "votes": 1, "replies": [] }, { "id": 1242410, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2021-03-17T15:41:30.037000", "content": "

That's a lot of models. 😁
\nCongrats. 🤘

", "votes": 1, "replies": [] }, { "id": 1242142, "author_name": "kalo", "author_url": "", "post_date": "2021-03-17T12:34:16.163000", "content": "

Congratulation! It's amazing how few commits there are!

", "votes": 1, "replies": [] }, { "id": 1241676, "author_name": "limzero", "author_url": "", "post_date": "2021-03-17T06:40:05.430000", "content": "

congratulation!

", "votes": 1, "replies": [] }, { "id": 1241665, "author_name": "Heroseo", "author_url": "", "post_date": "2021-03-17T06:30:36.883000", "content": "

Congrats on 1st place and great team effort!
\n@haqishen @boliu0 @garybios

\n

I'm looking forward to see your team solution
\nAnd thank you for always releasing minimal pipeline.

\n

p.s. I still remember the Melanoma competition. :)

", "votes": 1, "replies": [] }, { "id": 1241476, "author_name": "Miyatti", "author_url": "", "post_date": "2021-03-17T04:12:03.327000", "content": "

Congratulations on 1st place!! CV0.976 is very impressive.

", "votes": 1, "replies": [] }, { "id": 1241431, "author_name": "Reuben Schmidt", "author_url": "", "post_date": "2021-03-17T03:23:41.833000", "content": "

Congratulations. Fantastic effort

", "votes": 1, "replies": [] }, { "id": 1241426, "author_name": "RAHUL SINGH INDA", "author_url": "", "post_date": "2021-03-17T03:20:23.103000", "content": "

Congratulations on the win @haqishen

", "votes": 1, "replies": [] }, { "id": 1241360, "author_name": "KhanhVD", "author_url": "", "post_date": "2021-03-17T02:33:35.210000", "content": "

Strong Ensemble and solid CV. Well done congrats @haqishen @garybios and @boliu0. Looking forward to reading detail solution

", "votes": 1, "replies": [] }, { "id": 1241344, "author_name": "gao-hongnan", "author_url": "", "post_date": "2021-03-17T02:19:16.110000", "content": "

Too strong!!! Congrats!!!!!

", "votes": 1, "replies": [] }, { "id": 1241340, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-03-17T02:15:00.650000", "content": "

We had a CV of 0.9767 but only scored 0.971 at LB. i think we had a data leak

", "votes": 1, "replies": [ { "id": 1241348, "author_name": "Qishen Ha", "author_url": "", "post_date": "2021-03-17T02:21:12.187000", "content": "

You used pretrained weights from the forum right?

", "votes": 2, "replies": [] }, { "id": 1241353, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-03-17T02:27:45.663000", "content": "

yes . but scored 0.972 at private which we combined with our best ensemble to make our select submission

", "votes": 0, "replies": [] }, { "id": 1241365, "author_name": "Qishen Ha", "author_url": "", "post_date": "2021-03-17T02:35:27.693000", "content": "

When use such pretrained weights you should be cafeful on what data have been used to train it.

\n

That says, If sample A is trained in the pretrained weight, data leak will occur if sample A is validated in your finetune process.

", "votes": 4, "replies": [] }, { "id": 1241377, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-03-17T02:38:52.723000", "content": "

When i recheck our submission . 0.9767 was a CV of single model without using any of the pretrained weight of the forum.

", "votes": 1, "replies": [] }, { "id": 1241391, "author_name": "", "author_url": "", "post_date": "2021-03-17T02:50:02.240000", "content": "", "votes": 0, "replies": [] }, { "id": 1241390, "author_name": "Qishen Ha", "author_url": "", "post_date": "2021-03-17T02:50:02.240000", "content": "

Hope you can find out what caused this gap… 😂

", "votes": 4, "replies": [] }, { "id": 1241439, "author_name": "DatNT", "author_url": "", "post_date": "2021-03-17T03:30:59.420000", "content": "

i also think there are leaks on your cv. my local cv scores 0.974 and private lb scores 0.975

", "votes": 1, "replies": [] } ] }, { "id": 1246766, "author_name": "kcy4", "author_url": "", "post_date": "2021-03-21T04:18:37.680000", "content": "

Huge Congratulations!👍

", "votes": 2, "replies": [] }, { "id": 1241777, "author_name": "Theo Viel", "author_url": "", "post_date": "2021-03-17T07:38:03.333000", "content": "

Congratz ! An impressive result once again. I'll definitely study your solution carefully :)

", "votes": 2, "replies": [] }, { "id": 1241375, "author_name": "yukiya", "author_url": "", "post_date": "2021-03-17T02:38:22.340000", "content": "
\n

5 segmentation models (1024x1024 ~ 1536x1536)

\n
\n

What kind of hardware do you use to work with this size?
\nCongratulations on the win !

", "votes": 2, "replies": [ { "id": 1241397, "author_name": "Qishen Ha", "author_url": "", "post_date": "2021-03-17T02:58:12.280000", "content": "

Thanks!

\n

A month ago, I trained Unet-B5 1024x1024 on a single NVIDIA Quadro RTX6000 with 24GB and combine it with a B3 512 classification model to get CV 0.969 (5 fold seg + 5 fold cls ensemble gets pub and pvt LB 0.972)

\n

After that we didn't submit for a long time.

\n

You didn't see us on the LB before because we only predicted about 90% of the public test data at that time.

\n

I'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.

", "votes": 8, "replies": [] }, { "id": 1242639, "author_name": "Yassine Alouini", "author_url": "", "post_date": "2021-03-17T18:15:47.510000", "content": "

That would be very interesting, thanks in advance:

\n
\n

I'm cleaning our code and will soon release some kernels to show how we can train segmentation models using 1024 img on 16GB GPU as well.

\n
", "votes": 3, "replies": [] } ] }, { "id": 1241374, "author_name": "DungNB", "author_url": "", "post_date": "2021-03-17T02:37:57.783000", "content": "

congratulation!

", "votes": 2, "replies": [] }, { "id": 1241321, "author_name": "Mohammed Rizin V K", "author_url": "", "post_date": "2021-03-17T02:01:46.673000", "content": "

Congratz !!!!

\n
\n

which took 8.7h to run.

\n
\n

just too close to timeup

", "votes": 2, "replies": [] }, { "id": 1241337, "author_name": "Tensor Girl", "author_url": "", "post_date": "2021-03-17T02:12:18.373000", "content": "

@haqishen Congratulations on 1 st place Finish with too few submissions

", "votes": 1, "replies": [] }, { "id": 1248949, "author_name": "Sarthak Bhatt", "author_url": "", "post_date": "2021-03-23T00:41:49.170000", "content": "

Congratulations!🎉🎉

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1241316": "Congratulations to all winners and thanks to the kaggle, the host and my wonderful teammates! @boliu0 @garybios\n\nOur winning submission is based on 5 segmentation models (1024x1024 ~ 1536x1536) and 67 classification models (384x384 ~ 512x512), which took 8.7h to run.\n\nAmong our 22 submissions, this one has the 3rd highest pvt score but with the highest local cv —— **0.976057**\n\nThe cv strategy we use is Stratified Group K Fold.\n\nBrief summary of our solution has been released by @boliu0 :\nhttps://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/226633\n\nFurthermore, I've released our minimal pipeline here: https://www.kaggle.com/c/ranzcr-clip-catheter-line-classification/discussion/227128 which is able to reach **pub/pvt LB 0.972+** by training only on a single 16GB GPU.", "1241470": "Congrats )) great work. Well I think there is no need anymore for classification competitions they can put you always at the top and end the competition XD XD", "1241459": "Amazing, How can you train 67 classification model? By automatic searching different hyperparameter or purely by hand adjust?", "1241409": "congratulation, it's too crazy, 67 classification models and so big input segment model!!!", "1241339": "congratulation!", "1241333": "Congratulations!🎉🎉🎉 \nIt's very impressive to win a competition with only 22 submissions. Even the LB is stable.", "1241324": "> 5 segmentation models (1024x1024 ~ 1536~1536) and 67 classification models (384x384 ~ 512x512)\n\nmy god. but it's really creative you guys use big size for segmentation and only small size for classification. congrats!", "1241338": "congratulation!", "1244823": "Congratulations", "1244348": "Congrats ! Looking forward to your solution ", "1244318": "congratulation！！！", "1242410": "That's a lot of models. 😁 \nCongrats. 🤘 ", "1242142": "Congratulation! It's amazing how few commits there are!", "1241676": "congratulation!", "1241665": "Congrats on 1st place and great team effort!\n@haqishen @boliu0 @garybios \n\nI'm looking forward to see your team solution\nAnd thank you for always releasing minimal pipeline.\n\np.s. I still remember the Melanoma competition. :)", "1241476": "Congratulations on 1st place!! CV0.976 is very impressive.", "1241431": "Congratulations. Fantastic effort", "1241426": "Congratulations on the win @haqishen ", "1241360": "Strong Ensemble and solid CV. Well done congrats @haqishen @garybios and @boliu0. Looking forward to reading detail solution", "1241344": "Too strong!!! Congrats!!!!!", "1241340": "We had a CV of 0.9767 but only scored 0.971 at LB. i think we had a data leak", "1246766": "Huge Congratulations!👍", "1241777": "Congratz ! An impressive result once again. I'll definitely study your solution carefully :)", "1241375": "> 5 segmentation models (1024x1024 ~ 1536x1536)\n\nWhat kind of hardware do you use to work with this size?\nCongratulations on the win ! ", "1241374": "congratulation!", "1241321": "Congratz !!!! \n>which took 8.7h to run.\n\njust too close to timeup", "1241337": "@haqishen Congratulations on 1 st place Finish with too few submissions", "1248949": "Congratulations!🎉🎉" } }