{ "id": 264311, "title": "12th place solution", "url": "/competitions/siim-covid19-detection/writeups/inoichan-12th-place-solution", "author_name": "", "post_date": "2021-08-11T18:35:50.390Z", "votes": 21, "comment_count": 4, "views": 0, "content": "

Fisrst of all, many thanks to Kaggle and the hosts for hosting such an interesting competition, and congratulations to all the winners.

\n

I'm happy to have won a solo gold medal in this wonderful competition. I am grateful to Kaggler for teaming up with me so far. I have learned so much from them that I was able to win this gold medal. I also appreciate the Kaggle community. There is always a lot of knowledge being shared in Discussion and Code, and the top teams share their great solutions after the competition. Without it, I wouldn't have been able to grow as much as I am now. Thank you.

\n

Overview

\n

\"classification_detection_model

\n

Details are as follows.

\n

Study level and ‘none’

\n

In the first month, I had a very hard time getting any good CV score. After nearly 100 experiments, I found that data augmentation was not suitable, and when I changed the augmentation to the one shown by @haqishen in the RANZCR solution, the CV increased dramatically.
\nhttps://www.kaggle.com/haqishen/ranzcr-1st-place-soluiton-cls-model-small-ver

\n

CV was always better with 5-class classification including ‘none’ than with 4-class + ‘none’ separately. It was better to use sigmoid + bce than softmax + ce.

\n

Model architecture

\n

\n

Other tips;

\n\n

I use the aux loss that @hengck23 shared in the discussion.
\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240233

\n

Best single model of study level is EffnetV2L ATfL6, CV 0.4248 / LB 0.463 / PB 0.430 (LB/PB with ‘none 1 0 0 1 1’ in Image level).

\n

Swin transformer scored a little lower, but contributed to the ensemble.

\n

Image level

\n

In addition to Yolo, I used several mmdet models as shown in the overview picture.
\nBest model is Cascade RCNN Res2Net101, CV 0.0897 / LB 0.094 / PB 0.096.

\n

WBF (th=0.5) ensemble boosted score to LB 0.098 / PB 0.101.

\n

Detection confidence was calibrated as follow;

\n
det_conf = det_conf * (1 - none)**0.5\n
\n

Inference code

\n

https://www.kaggle.com/inoueu1/siim-sub145-lp-c2-effv2m-l-p-swin-mmdet?scriptVersionId=70843218

", "messages": [ { "id": "1466926", "postDate": "08/11/2021 17:26:32", "content": "

Fisrst of all, many thanks to Kaggle and the hosts for hosting such an interesting competition, and congratulations to all the winners.

\n

I'm happy to have won a solo gold medal in this wonderful competition. I am grateful to Kaggler for teaming up with me so far. I have learned so much from them that I was able to win this gold medal. I also appreciate the Kaggle community. There is always a lot of knowledge being shared in Discussion and Code, and the top teams share their great solutions after the competition. Without it, I wouldn't have been able to grow as much as I am now. Thank you.

\n

Overview

\n

\"classification_detection_model

\n

Details are as follows.

\n

Study level and ‘none’

\n

In the first month, I had a very hard time getting any good CV score. After nearly 100 experiments, I found that data augmentation was not suitable, and when I changed the augmentation to the one shown by @haqishen in the RANZCR solution, the CV increased dramatically.
\nhttps://www.kaggle.com/haqishen/ranzcr-1st-place-soluiton-cls-model-small-ver

\n

CV was always better with 5-class classification including ‘none’ than with 4-class + ‘none’ separately. It was better to use sigmoid + bce than softmax + ce.

\n

Model architecture

\n

\n

Other tips;

\n\n

I use the aux loss that @hengck23 shared in the discussion.
\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240233

\n

Best single model of study level is EffnetV2L ATfL6, CV 0.4248 / LB 0.463 / PB 0.430 (LB/PB with ‘none 1 0 0 1 1’ in Image level).

\n

Swin transformer scored a little lower, but contributed to the ensemble.

\n

Image level

\n

In addition to Yolo, I used several mmdet models as shown in the overview picture.
\nBest model is Cascade RCNN Res2Net101, CV 0.0897 / LB 0.094 / PB 0.096.

\n

WBF (th=0.5) ensemble boosted score to LB 0.098 / PB 0.101.

\n

Detection confidence was calibrated as follow;

\n
det_conf = det_conf * (1 - none)**0.5\n
\n

Inference code

\n

https://www.kaggle.com/inoueu1/siim-sub145-lp-c2-effv2m-l-p-swin-mmdet?scriptVersionId=70843218

", "rawMarkdown": "Fisrst of all, many thanks to Kaggle and the hosts for hosting such an interesting competition, and congratulations to all the winners. \n\nI'm happy to have won a solo gold medal in this wonderful competition. I am grateful to Kaggler for teaming up with me so far. I have learned so much from them that I was able to win this gold medal. I also appreciate the Kaggle community. There is always a lot of knowledge being shared in Discussion and Code, and the top teams share their great solutions after the competition. Without it, I wouldn't have been able to grow as much as I am now. Thank you.\n\n# Overview\n\n![classification_detection_model (1)](https://user-images.githubusercontent.com/37664066/129072710-1463016b-4ca6-410c-a38b-381caaf1838b.jpg)\n\nDetails are as follows.\n\n# Study level and ‘none’\n\n\nIn the first month, I had a very hard time getting any good CV score. After nearly 100 experiments, I found that data augmentation was not suitable, and when I changed the augmentation to the one shown by @haqishen in the RANZCR solution, the CV increased dramatically.\nhttps://www.kaggle.com/haqishen/ranzcr-1st-place-soluiton-cls-model-small-ver\n\nCV was always better with 5-class classification including ‘none’ than with 4-class + ‘none’ separately. It was better to use sigmoid + bce than softmax + ce.\n\n### Model architecture\n\n \n\n### Other tips;\n\n- Five fold cross validation split by Patient\n- model: Efficientnet v2 m, l and swin transformer base(384)\n- optimizer: AdamW\n- scheduler: Cosine Annealing (lr 1e-4 -> 1e-5)\n- epoch: 20 ~ 30\n- Additional segmentation task, bce + symmetric lovasz (1 : 1)\n- Aux loss : Cls loss = 1 : 1 (Aux loss is about twice as large as Cls loss)\n- Pseudo labeling (bimcv+, record)\n- TTA: HFlip\n\nI use the aux loss that @hengck23 shared in the discussion.\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240233\n\nBest single model of study level is EffnetV2L ATfL6, CV 0.4248 / LB 0.463 / PB 0.430 (LB/PB with ‘none 1 0 0 1 1’ in Image level).\n\nSwin transformer scored a little lower, but contributed to the ensemble.\n\n\n# Image level\n\nIn addition to Yolo, I used several mmdet models as shown in the overview picture.\nBest model is Cascade RCNN Res2Net101, CV 0.0897 / LB 0.094 / PB 0.096.\n\nWBF (th=0.5) ensemble boosted score to LB 0.098 / PB 0.101.\n\nDetection confidence was calibrated as follow;\n```\ndet_conf = det_conf * (1 - none)**0.5\n```\n \n \n## Inference code\n\nhttps://www.kaggle.com/inoueu1/siim-sub145-lp-c2-effv2m-l-p-swin-mmdet?scriptVersionId=70843218", "votes": null }, { "id": "1471418", "postDate": "08/14/2021 07:36:11", "content": "

Thanks for the great writeup! Your aux loss implementation is interesting. Also, congrats to your gold. I have one question, did you get a CV boost from your confidence calibration? We tried to make this work, but it did not help unfortunately.

", "rawMarkdown": "Thanks for the great writeup! Your aux loss implementation is interesting. Also, congrats to your gold. I have one question, did you get a CV boost from your confidence calibration? We tried to make this work, but it did not help unfortunately.", "votes": null }, { "id": "1471869", "postDate": "08/14/2021 14:23:08", "content": "

Thank you!!
\nI used only LB to see the effect of the confidence calibration. I used it for the final sub because it was said to boost scores in the previous VinBigData competition, and because LB had a slightly better each time. However, the effect is marginal, about at most 0.0005, I think.

", "rawMarkdown": "Thank you!!\nI used only LB to see the effect of the confidence calibration. I used it for the final sub because it was said to boost scores in the previous VinBigData competition, and because LB had a slightly better each time. However, the effect is marginal, about at most 0.0005, I think.", "votes": null }, { "id": "1480134", "postDate": "08/18/2021 20:17:58", "content": "

Ok interesting, thank you! :)

", "rawMarkdown": "Ok interesting, thank you! :)", "votes": null }, { "id": "1610363", "postDate": "12/07/2021 06:37:58", "content": "

Thank you for sharing! I think it is a great solution. I have one question.
\n1.How is this kind of model structure created?

", "rawMarkdown": "Thank you for sharing! I think it is a great solution. I have one question.\n1.How is this kind of model structure created?", "votes": null } ], "comments": [ { "id": 1471418, "author_name": "nicohrubec", "author_url": "", "post_date": "08/14/2021 07:36:11", "content": "

Thanks for the great writeup! Your aux loss implementation is interesting. Also, congrats to your gold. I have one question, did you get a CV boost from your confidence calibration? We tried to make this work, but it did not help unfortunately.

", "votes": null, "replies": [ { "id": 1471869, "author_name": "inoueu1", "author_url": "", "post_date": "08/14/2021 14:23:08", "content": "

Thank you!!
\nI used only LB to see the effect of the confidence calibration. I used it for the final sub because it was said to boost scores in the previous VinBigData competition, and because LB had a slightly better each time. However, the effect is marginal, about at most 0.0005, I think.

", "votes": null, "replies": [] }, { "id": 1480134, "author_name": "nicohrubec", "author_url": "", "post_date": "08/18/2021 20:17:58", "content": "

Ok interesting, thank you! :)

", "votes": null, "replies": [] } ] }, { "id": 1610363, "author_name": "kotashimomura", "author_url": "", "post_date": "12/07/2021 06:37:58", "content": "

Thank you for sharing! I think it is a great solution. I have one question.
\n1.How is this kind of model structure created?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1466926": "Fisrst of all, many thanks to Kaggle and the hosts for hosting such an interesting competition, and congratulations to all the winners. \n\nI'm happy to have won a solo gold medal in this wonderful competition. I am grateful to Kaggler for teaming up with me so far. I have learned so much from them that I was able to win this gold medal. I also appreciate the Kaggle community. There is always a lot of knowledge being shared in Discussion and Code, and the top teams share their great solutions after the competition. Without it, I wouldn't have been able to grow as much as I am now. Thank you.\n\n# Overview\n\n![classification_detection_model (1)](https://user-images.githubusercontent.com/37664066/129072710-1463016b-4ca6-410c-a38b-381caaf1838b.jpg)\n\nDetails are as follows.\n\n# Study level and ‘none’\n\n\nIn the first month, I had a very hard time getting any good CV score. After nearly 100 experiments, I found that data augmentation was not suitable, and when I changed the augmentation to the one shown by @haqishen in the RANZCR solution, the CV increased dramatically.\nhttps://www.kaggle.com/haqishen/ranzcr-1st-place-soluiton-cls-model-small-ver\n\nCV was always better with 5-class classification including ‘none’ than with 4-class + ‘none’ separately. It was better to use sigmoid + bce than softmax + ce.\n\n### Model architecture\n\n \n\n### Other tips;\n\n- Five fold cross validation split by Patient\n- model: Efficientnet v2 m, l and swin transformer base(384)\n- optimizer: AdamW\n- scheduler: Cosine Annealing (lr 1e-4 -> 1e-5)\n- epoch: 20 ~ 30\n- Additional segmentation task, bce + symmetric lovasz (1 : 1)\n- Aux loss : Cls loss = 1 : 1 (Aux loss is about twice as large as Cls loss)\n- Pseudo labeling (bimcv+, record)\n- TTA: HFlip\n\nI use the aux loss that @hengck23 shared in the discussion.\nhttps://www.kaggle.com/c/siim-covid19-detection/discussion/240233\n\nBest single model of study level is EffnetV2L ATfL6, CV 0.4248 / LB 0.463 / PB 0.430 (LB/PB with ‘none 1 0 0 1 1’ in Image level).\n\nSwin transformer scored a little lower, but contributed to the ensemble.\n\n\n# Image level\n\nIn addition to Yolo, I used several mmdet models as shown in the overview picture.\nBest model is Cascade RCNN Res2Net101, CV 0.0897 / LB 0.094 / PB 0.096.\n\nWBF (th=0.5) ensemble boosted score to LB 0.098 / PB 0.101.\n\nDetection confidence was calibrated as follow;\n```\ndet_conf = det_conf * (1 - none)**0.5\n```\n \n \n## Inference code\n\nhttps://www.kaggle.com/inoueu1/siim-sub145-lp-c2-effv2m-l-p-swin-mmdet?scriptVersionId=70843218", "1471418": "Thanks for the great writeup! Your aux loss implementation is interesting. Also, congrats to your gold. I have one question, did you get a CV boost from your confidence calibration? We tried to make this work, but it did not help unfortunately.", "1471869": "Thank you!!\nI used only LB to see the effect of the confidence calibration. I used it for the final sub because it was said to boost scores in the previous VinBigData competition, and because LB had a slightly better each time. However, the effect is marginal, about at most 0.0005, I think.", "1480134": "Ok interesting, thank you! :)", "1610363": "Thank you for sharing! I think it is a great solution. I have one question.\n1.How is this kind of model structure created?" }, "source": "meta" }