{ "id": 263683, "title": "10th place with code", "url": "/competitions/siim-covid19-detection/writeups/guanshuo-xu-10th-place-with-code", "author_name": "", "post_date": "2021-08-24T03:45:07.463Z", "votes": 55, "comment_count": 10, "views": 0, "content": "

Training code: https://github.com/GuanshuoXu/SIIM-FISABIO-RSNA-COVID-19-Detection-10th-place-solution
\nInference kernel: https://www.kaggle.com/wowfattie/notebook68483076ff?scriptVersionId=70714056
\nvideo: https://drive.google.com/file/d/15l82vX5TCD-ltcmxAcOJ0A-c5vYOcY3Z/view?usp=sharing

\n

My solution is comprised of a detection model to predict all the targets and some expert models for performance enhancement.

\n

Since there is a detection target, it's more convenient to work with a detection model. To predict all the six targets, the whole image was used as bbox for the four study level labels, 'none' target was not assigned any bbox and was predicted by multiplying the score of the top 3 opacity bbox predictions of each image. The idea was mainly inspired from the Vinbigdata top solutions. My 5-fold efficientdet-D5 trained with this method achieved public LB 0.622 and private LB 0.618.

\n

To improve performance, I added some yolov5 models for opacity detection, efficient-B7&B8 for none prediction, and efficient-B7&B8 for study-level label predictions. I used RICCORD, BIMCV+ and BIMCV- dataset for pseudo-labeling. Blending them all gave me public LB 0.639 and private LB 0.624.

", "messages": [ { "id": "1462806", "postDate": "08/10/2021 02:15:33", "content": "

Training code: https://github.com/GuanshuoXu/SIIM-FISABIO-RSNA-COVID-19-Detection-10th-place-solution
\nInference kernel: https://www.kaggle.com/wowfattie/notebook68483076ff?scriptVersionId=70714056
\nvideo: https://drive.google.com/file/d/15l82vX5TCD-ltcmxAcOJ0A-c5vYOcY3Z/view?usp=sharing

\n

My solution is comprised of a detection model to predict all the targets and some expert models for performance enhancement.

\n

Since there is a detection target, it's more convenient to work with a detection model. To predict all the six targets, the whole image was used as bbox for the four study level labels, 'none' target was not assigned any bbox and was predicted by multiplying the score of the top 3 opacity bbox predictions of each image. The idea was mainly inspired from the Vinbigdata top solutions. My 5-fold efficientdet-D5 trained with this method achieved public LB 0.622 and private LB 0.618.

\n

To improve performance, I added some yolov5 models for opacity detection, efficient-B7&B8 for none prediction, and efficient-B7&B8 for study-level label predictions. I used RICCORD, BIMCV+ and BIMCV- dataset for pseudo-labeling. Blending them all gave me public LB 0.639 and private LB 0.624.

", "rawMarkdown": "Training code: https://github.com/GuanshuoXu/SIIM-FISABIO-RSNA-COVID-19-Detection-10th-place-solution\nInference kernel: https://www.kaggle.com/wowfattie/notebook68483076ff?scriptVersionId=70714056\nvideo: https://drive.google.com/file/d/15l82vX5TCD-ltcmxAcOJ0A-c5vYOcY3Z/view?usp=sharing\n\nMy solution is comprised of a detection model to predict all the targets and some expert models for performance enhancement.\n\nSince there is a detection target, it's more convenient to work with a detection model. To predict all the six targets, the whole image was used as bbox for the four study level labels, 'none' target was not assigned any bbox and was predicted by multiplying the score of the top 3 opacity bbox predictions of each image. The idea was mainly inspired from the Vinbigdata top solutions. My 5-fold efficientdet-D5 trained with this method achieved public LB 0.622 and private LB 0.618. \n\nTo improve performance, I added some yolov5 models for opacity detection, efficient-B7&B8 for none prediction, and efficient-B7&B8 for study-level label predictions. I used RICCORD, BIMCV+ and BIMCV- dataset for pseudo-labeling. Blending them all gave me public LB 0.639 and private LB 0.624.", "votes": null }, { "id": "1463665", "postDate": "08/10/2021 09:21:00", "content": "

Great works

", "rawMarkdown": "Great works", "votes": null }, { "id": "1463755", "postDate": "08/10/2021 10:13:55", "content": "
\n

to predict all the six targets, the whole image was used as bbox for the four study level labels,

\n
\n

I thought of that only two days before deadline and couldn't include the study predictions in time. SIgh.

", "rawMarkdown": "> to predict all the six targets, the whole image was used as bbox for the four study level labels, \n\nI thought of that only two days before deadline and couldn't include the study predictions in time. SIgh.", "votes": null }, { "id": "1464947", "postDate": "08/10/2021 19:16:18", "content": "

Congratulations on your gold medal! 🎉🎉🎉
\nWe also tried training EfficientDet-D7 with 6 targets but weren't able to get good performance for study level predictions. Did you do anything special in your training? It would be awesome if you could share your augmentations and some hyperparameters. Thanks a lot!

", "rawMarkdown": "Congratulations on your gold medal! 🎉🎉🎉\nWe also tried training EfficientDet-D7 with 6 targets but weren't able to get good performance for study level predictions. Did you do anything special in your training? It would be awesome if you could share your augmentations and some hyperparameters. Thanks a lot!", "votes": null }, { "id": "1465191", "postDate": "08/10/2021 23:29:16", "content": "

I will publish the code in a few days

", "rawMarkdown": "I will publish the code in a few days", "votes": null }, { "id": "1465955", "postDate": "08/11/2021 08:42:59", "content": "

Congrats for your amazing solo performance (counting only ~20 subs)!!

\n
\n

Blending them all gave me public LB 0.639 and private LB 0.624

\n
\n

What bout the PVT score your highest Public LB 0.647? was from same blend of models ?

", "rawMarkdown": "Congrats for your amazing solo performance (counting only ~20 subs)!! \n\n>Blending them all gave me public LB 0.639 and private LB 0.624\n\nWhat bout the PVT score your highest Public LB 0.647? was from same blend of models ?", "votes": null }, { "id": "1466249", "postDate": "08/11/2021 11:30:27", "content": "

The 0.647 was a blend with the predictions using tabular data (annotations) in RICORD and BIMCV as features. It worked pretty well in public LB but somehow failed in private LB.

", "rawMarkdown": "The 0.647 was a blend with the predictions using tabular data (annotations) in RICORD and BIMCV as features. It worked pretty well in public LB but somehow failed in private LB.", "votes": null }, { "id": "1466868", "postDate": "08/11/2021 17:00:01", "content": "

Thank you!

", "rawMarkdown": "Thank you!", "votes": null }, { "id": "1467325", "postDate": "08/11/2021 23:42:43", "content": "

Good luck for you next time bro, keep it up ! 👌

", "rawMarkdown": "Good luck for you next time bro, keep it up ! 👌", "votes": null }, { "id": "1467703", "postDate": "08/12/2021 05:39:12", "content": "

A real gold digger. Congrats @wowfattie

", "rawMarkdown": "A real gold digger. Congrats @wowfattie", "votes": null }, { "id": "1477685", "postDate": "08/17/2021 15:29:16", "content": "

Hi, do you use the efficiendets from here https://github.com/rwightman/efficientdet-pytorch?

", "rawMarkdown": "Hi, do you use the efficiendets from here https://github.com/rwightman/efficientdet-pytorch?", "votes": null } ], "comments": [ { "id": 1463665, "author_name": "givkashi", "author_url": "", "post_date": "08/10/2021 09:21:00", "content": "

Great works

", "votes": null, "replies": [] }, { "id": 1463755, "author_name": "cpmpml", "author_url": "", "post_date": "08/10/2021 10:13:55", "content": "
\n

to predict all the six targets, the whole image was used as bbox for the four study level labels,

\n
\n

I thought of that only two days before deadline and couldn't include the study predictions in time. SIgh.

", "votes": null, "replies": [] }, { "id": 1464947, "author_name": "kokkini", "author_url": "", "post_date": "08/10/2021 19:16:18", "content": "

Congratulations on your gold medal! 🎉🎉🎉
\nWe also tried training EfficientDet-D7 with 6 targets but weren't able to get good performance for study level predictions. Did you do anything special in your training? It would be awesome if you could share your augmentations and some hyperparameters. Thanks a lot!

", "votes": null, "replies": [ { "id": 1465191, "author_name": "wowfattie", "author_url": "", "post_date": "08/10/2021 23:29:16", "content": "

I will publish the code in a few days

", "votes": null, "replies": [] }, { "id": 1466868, "author_name": "kokkini", "author_url": "", "post_date": "08/11/2021 17:00:01", "content": "

Thank you!

", "votes": null, "replies": [] } ] }, { "id": 1465955, "author_name": "imeintanis", "author_url": "", "post_date": "08/11/2021 08:42:59", "content": "

Congrats for your amazing solo performance (counting only ~20 subs)!!

\n
\n

Blending them all gave me public LB 0.639 and private LB 0.624

\n
\n

What bout the PVT score your highest Public LB 0.647? was from same blend of models ?

", "votes": null, "replies": [ { "id": 1466249, "author_name": "wowfattie", "author_url": "", "post_date": "08/11/2021 11:30:27", "content": "

The 0.647 was a blend with the predictions using tabular data (annotations) in RICORD and BIMCV as features. It worked pretty well in public LB but somehow failed in private LB.

", "votes": null, "replies": [] } ] }, { "id": 1467325, "author_name": "iniestamoh", "author_url": "", "post_date": "08/11/2021 23:42:43", "content": "

Good luck for you next time bro, keep it up ! 👌

", "votes": null, "replies": [] }, { "id": 1467703, "author_name": "kingabzpro", "author_url": "", "post_date": "08/12/2021 05:39:12", "content": "

A real gold digger. Congrats @wowfattie

", "votes": null, "replies": [] }, { "id": 1477685, "author_name": "jackchungchiehyu", "author_url": "", "post_date": "08/17/2021 15:29:16", "content": "

Hi, do you use the efficiendets from here https://github.com/rwightman/efficientdet-pytorch?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1462806": "Training code: https://github.com/GuanshuoXu/SIIM-FISABIO-RSNA-COVID-19-Detection-10th-place-solution\nInference kernel: https://www.kaggle.com/wowfattie/notebook68483076ff?scriptVersionId=70714056\nvideo: https://drive.google.com/file/d/15l82vX5TCD-ltcmxAcOJ0A-c5vYOcY3Z/view?usp=sharing\n\nMy solution is comprised of a detection model to predict all the targets and some expert models for performance enhancement.\n\nSince there is a detection target, it's more convenient to work with a detection model. To predict all the six targets, the whole image was used as bbox for the four study level labels, 'none' target was not assigned any bbox and was predicted by multiplying the score of the top 3 opacity bbox predictions of each image. The idea was mainly inspired from the Vinbigdata top solutions. My 5-fold efficientdet-D5 trained with this method achieved public LB 0.622 and private LB 0.618. \n\nTo improve performance, I added some yolov5 models for opacity detection, efficient-B7&B8 for none prediction, and efficient-B7&B8 for study-level label predictions. I used RICCORD, BIMCV+ and BIMCV- dataset for pseudo-labeling. Blending them all gave me public LB 0.639 and private LB 0.624.", "1463665": "Great works", "1463755": "> to predict all the six targets, the whole image was used as bbox for the four study level labels, \n\nI thought of that only two days before deadline and couldn't include the study predictions in time. SIgh.", "1464947": "Congratulations on your gold medal! 🎉🎉🎉\nWe also tried training EfficientDet-D7 with 6 targets but weren't able to get good performance for study level predictions. Did you do anything special in your training? It would be awesome if you could share your augmentations and some hyperparameters. Thanks a lot!", "1465191": "I will publish the code in a few days", "1465955": "Congrats for your amazing solo performance (counting only ~20 subs)!! \n\n>Blending them all gave me public LB 0.639 and private LB 0.624\n\nWhat bout the PVT score your highest Public LB 0.647? was from same blend of models ?", "1466249": "The 0.647 was a blend with the predictions using tabular data (annotations) in RICORD and BIMCV as features. It worked pretty well in public LB but somehow failed in private LB.", "1466868": "Thank you!", "1467325": "Good luck for you next time bro, keep it up ! 👌", "1467703": "A real gold digger. Congrats @wowfattie", "1477685": "Hi, do you use the efficiendets from here https://github.com/rwightman/efficientdet-pytorch?" }, "source": "meta" }