{ "id": 263748, "title": "7th place solution with Code", "url": "/competitions/siim-covid19-detection/writeups/dsmlkz-school-zerde-7th-place-solution-with-code", "author_name": "", "post_date": "2021-08-15T20:15:46.153Z", "votes": 26, "comment_count": 7, "views": 0, "content": "

First, thanks for organizing this interesting competition and congrats to all winners!

\n

Validation:

\n

We used iterative-stratification with 5 folds (https://github.com/trent-b/iterative-stratification) stratified by study level classes and number of boxes.

\n

Study level:

\n

We used efficientnet-b7, v2s, v2m, and v2l with aux branches after 3 different blocks. The models were trained on 3 folds and on different image resolutions (512, 640, 768) to produce 14 classifiers.

\n

We used simple averaging for ensembling the models. LB mAP for study level was ~41.5-41.6

\n

Image level:

\n

We used mmdetection library to train detectoRS50, universeNet50, and universeNet101. detectoRS50 and universeNet50 were trained on one fold, and universeNet101 was trained on each fold + pseudo labels for public data using the universeNet50 model.

\n

WBF did not work for us, so we decided to use NMW for ensembling from https://github.com/ZFTurbo/Weighted-Boxes-Fusion.

\n

TTA: HorizontalFlip for detectoRS and multi-scale TTA for all universeNet models on [(640, 640), (800, 800)].

\n

Binary classifiers were trained in the same manner as study level models, 3 fold ensemble was used.

\n

Augmentations:

\n

Our augmentations include HorizontalFlip, RandomCrop (for study level), ShiftScaleRotate, CLAHE, RandomGamma, Cutout from albumentations library (https://albumentations.ai/ ).

\n

UPD: Code available at https://github.com/AidynUbingazhibov/SIIM-FISABIO-RSNA-COVID-19-Detection

\n

Inference kernel: https://www.kaggle.com/aidynub/validation-pipeline-4cls-10mdlswp-bincl-3mdls-d?scriptVersionId=70822471

", "messages": [ { "id": "1463236", "postDate": "08/10/2021 06:22:41", "content": "

First, thanks for organizing this interesting competition and congrats to all winners!

\n

Validation:

\n

We used iterative-stratification with 5 folds (https://github.com/trent-b/iterative-stratification) stratified by study level classes and number of boxes.

\n

Study level:

\n

We used efficientnet-b7, v2s, v2m, and v2l with aux branches after 3 different blocks. The models were trained on 3 folds and on different image resolutions (512, 640, 768) to produce 14 classifiers.

\n

We used simple averaging for ensembling the models. LB mAP for study level was ~41.5-41.6

\n

Image level:

\n

We used mmdetection library to train detectoRS50, universeNet50, and universeNet101. detectoRS50 and universeNet50 were trained on one fold, and universeNet101 was trained on each fold + pseudo labels for public data using the universeNet50 model.

\n

WBF did not work for us, so we decided to use NMW for ensembling from https://github.com/ZFTurbo/Weighted-Boxes-Fusion.

\n

TTA: HorizontalFlip for detectoRS and multi-scale TTA for all universeNet models on [(640, 640), (800, 800)].

\n

Binary classifiers were trained in the same manner as study level models, 3 fold ensemble was used.

\n

Augmentations:

\n

Our augmentations include HorizontalFlip, RandomCrop (for study level), ShiftScaleRotate, CLAHE, RandomGamma, Cutout from albumentations library (https://albumentations.ai/ ).

\n

UPD: Code available at https://github.com/AidynUbingazhibov/SIIM-FISABIO-RSNA-COVID-19-Detection

\n

Inference kernel: https://www.kaggle.com/aidynub/validation-pipeline-4cls-10mdlswp-bincl-3mdls-d?scriptVersionId=70822471

", "rawMarkdown": "First, thanks for organizing this interesting competition and congrats to all winners!\n\n**Validation:**\n\nWe used iterative-stratification with 5 folds (https://github.com/trent-b/iterative-stratification) stratified by study level classes and number of boxes.\n\n**Study level:**\n\nWe used efficientnet-b7, v2s, v2m, and v2l with aux branches after 3 different blocks. The models were trained on 3 folds and on different image resolutions (512, 640, 768) to produce 14 classifiers. \n\nWe used simple averaging for ensembling the models. LB mAP for study level was ~41.5-41.6\n\n\n\n**Image level:**\n\nWe used mmdetection library to train detectoRS50, universeNet50, and universeNet101. detectoRS50 and universeNet50 were trained on one fold, and universeNet101 was trained on each fold + pseudo labels for public data using the universeNet50 model.\n\nWBF did not work for us, so we decided to use NMW for ensembling from https://github.com/ZFTurbo/Weighted-Boxes-Fusion.\n\nTTA: HorizontalFlip for detectoRS and multi-scale TTA for all universeNet models on [(640, 640), (800, 800)].\n\nBinary classifiers were trained in the same manner as study level models, 3 fold ensemble was used.\n\n**Augmentations:**\n\nOur augmentations include HorizontalFlip, RandomCrop (for study level), ShiftScaleRotate, CLAHE, RandomGamma, Cutout from albumentations library (https://albumentations.ai/ ).\n\nUPD: Code available at https://github.com/AidynUbingazhibov/SIIM-FISABIO-RSNA-COVID-19-Detection\n\nInference kernel: https://www.kaggle.com/aidynub/validation-pipeline-4cls-10mdlswp-bincl-3mdls-d?scriptVersionId=70822471", "votes": null }, { "id": "1463590", "postDate": "08/10/2021 08:46:00", "content": "

congrat for the result an hard work. It would be nice if I have those questions answered by you:

\n
    \n
  1. You have 4 types of model, 3 folds, 3 image resolution, so why 14 classifiers ?
    2. \n
  2. Your study level on reached 41.5, 41.6. Is it including prediction none 1 0 0 1 1 for all image level? May I know the private score ?
    3. \n
", "rawMarkdown": "congrat for the result an hard work. It would be nice if I have those questions answered by you:\n1. You have 4 types of model, 3 folds, 3 image resolution, so why 14 classifiers ?\n2. Your study level on reached 41.5, 41.6. Is it including prediction none 1 0 0 1 1 for all image level? May I know the private score ?", "votes": null }, { "id": "1463666", "postDate": "08/10/2021 09:21:16", "content": "

congratulations

", "rawMarkdown": "congratulations", "votes": null }, { "id": "1463713", "postDate": "08/10/2021 09:43:58", "content": "

Thanks!

\n
    \n

  1. We haven't tried all the combinations, it was rather sporadic. From all of our experiments we chose 14 that had decent validation mAP.

    2. \n

  2. Not including. I don't have the exact private score for that since we didn't submit all 14 models separately. We have ensemble of 8 models that gave us 0.414 on public, and 0.392 on private.

    3. \n
", "rawMarkdown": "Thanks! \n\n1. We haven't tried all the combinations, it was rather sporadic. From all of our experiments we chose 14 that had decent validation mAP.\n\n2. Not including. I don't have the exact private score for that since we didn't submit all 14 models separately. We have ensemble of 8 models that gave us 0.414 on public, and 0.392 on private.", "votes": null }, { "id": "1464139", "postDate": "08/10/2021 13:11:32", "content": "

Thanks for sharing the solution and congratulations, mate.

", "rawMarkdown": "Thanks for sharing the solution and congratulations, mate.", "votes": null }, { "id": "1464201", "postDate": "08/10/2021 13:34:18", "content": "

Congrats!! seems your study level models ensemble were quite strong!! good job

\n
    \n

  1. do you know the CV, LB scores of your OD part ?

    2. \n

  2. any external data used ?

    3. \n
", "rawMarkdown": "Congrats!! seems your study level models ensemble were quite strong!! good job \n1. do you know the CV, LB scores of your OD part ? \n\n2. any external data used ?", "votes": null }, { "id": "1464241", "postDate": "08/10/2021 13:55:19", "content": "

Thank you!

\n
    \n

  1. 5 folds UniverseNet101: 57.7 / 57 / 54.8 / 54 / 53 mAP (images without bboxes are kept). LB: 0.096 / 0.096 / 0.094 / 0.096 / 0.096.

    2. \n

  2. No external dataset.

    3. \n
", "rawMarkdown": "Thank you!\n\n1. 5 folds UniverseNet101: 57.7 / 57 / 54.8 / 54 / 53 mAP (images without bboxes are kept). LB: 0.096 / 0.096 / 0.094 / 0.096 / 0.096.\n\n2. No external dataset.", "votes": null }, { "id": "1464369", "postDate": "08/10/2021 14:47:05", "content": "

congrats @aidynub 🎉🎉

", "rawMarkdown": "congrats @aidynub 🎉🎉", "votes": null } ], "comments": [ { "id": 1463590, "author_name": "namgalielei", "author_url": "", "post_date": "08/10/2021 08:46:00", "content": "

congrat for the result an hard work. It would be nice if I have those questions answered by you:

\n
    \n
  1. You have 4 types of model, 3 folds, 3 image resolution, so why 14 classifiers ?
    2. \n
  2. Your study level on reached 41.5, 41.6. Is it including prediction none 1 0 0 1 1 for all image level? May I know the private score ?
    3. \n
", "votes": null, "replies": [ { "id": 1463713, "author_name": "aidynub", "author_url": "", "post_date": "08/10/2021 09:43:58", "content": "

Thanks!

\n
    \n

  1. We haven't tried all the combinations, it was rather sporadic. From all of our experiments we chose 14 that had decent validation mAP.

    2. \n

  2. Not including. I don't have the exact private score for that since we didn't submit all 14 models separately. We have ensemble of 8 models that gave us 0.414 on public, and 0.392 on private.

    3. \n
", "votes": null, "replies": [] } ] }, { "id": 1463666, "author_name": "givkashi", "author_url": "", "post_date": "08/10/2021 09:21:16", "content": "

congratulations

", "votes": null, "replies": [] }, { "id": 1464139, "author_name": "furcifer", "author_url": "", "post_date": "08/10/2021 13:11:32", "content": "

Thanks for sharing the solution and congratulations, mate.

", "votes": null, "replies": [] }, { "id": 1464201, "author_name": "imeintanis", "author_url": "", "post_date": "08/10/2021 13:34:18", "content": "

Congrats!! seems your study level models ensemble were quite strong!! good job

\n
    \n

  1. do you know the CV, LB scores of your OD part ?

    2. \n

  2. any external data used ?

    3. \n
", "votes": null, "replies": [ { "id": 1464241, "author_name": "aidynub", "author_url": "", "post_date": "08/10/2021 13:55:19", "content": "

Thank you!

\n
    \n

  1. 5 folds UniverseNet101: 57.7 / 57 / 54.8 / 54 / 53 mAP (images without bboxes are kept). LB: 0.096 / 0.096 / 0.094 / 0.096 / 0.096.

    2. \n

  2. No external dataset.

    3. \n
", "votes": null, "replies": [] } ] }, { "id": 1464369, "author_name": "jarupula", "author_url": "", "post_date": "08/10/2021 14:47:05", "content": "

congrats @aidynub 🎉🎉

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1463236": "First, thanks for organizing this interesting competition and congrats to all winners!\n\n**Validation:**\n\nWe used iterative-stratification with 5 folds (https://github.com/trent-b/iterative-stratification) stratified by study level classes and number of boxes.\n\n**Study level:**\n\nWe used efficientnet-b7, v2s, v2m, and v2l with aux branches after 3 different blocks. The models were trained on 3 folds and on different image resolutions (512, 640, 768) to produce 14 classifiers. \n\nWe used simple averaging for ensembling the models. LB mAP for study level was ~41.5-41.6\n\n\n\n**Image level:**\n\nWe used mmdetection library to train detectoRS50, universeNet50, and universeNet101. detectoRS50 and universeNet50 were trained on one fold, and universeNet101 was trained on each fold + pseudo labels for public data using the universeNet50 model.\n\nWBF did not work for us, so we decided to use NMW for ensembling from https://github.com/ZFTurbo/Weighted-Boxes-Fusion.\n\nTTA: HorizontalFlip for detectoRS and multi-scale TTA for all universeNet models on [(640, 640), (800, 800)].\n\nBinary classifiers were trained in the same manner as study level models, 3 fold ensemble was used.\n\n**Augmentations:**\n\nOur augmentations include HorizontalFlip, RandomCrop (for study level), ShiftScaleRotate, CLAHE, RandomGamma, Cutout from albumentations library (https://albumentations.ai/ ).\n\nUPD: Code available at https://github.com/AidynUbingazhibov/SIIM-FISABIO-RSNA-COVID-19-Detection\n\nInference kernel: https://www.kaggle.com/aidynub/validation-pipeline-4cls-10mdlswp-bincl-3mdls-d?scriptVersionId=70822471", "1463590": "congrat for the result an hard work. It would be nice if I have those questions answered by you:\n1. You have 4 types of model, 3 folds, 3 image resolution, so why 14 classifiers ?\n2. Your study level on reached 41.5, 41.6. Is it including prediction none 1 0 0 1 1 for all image level? May I know the private score ?", "1463666": "congratulations", "1463713": "Thanks! \n\n1. We haven't tried all the combinations, it was rather sporadic. From all of our experiments we chose 14 that had decent validation mAP.\n\n2. Not including. I don't have the exact private score for that since we didn't submit all 14 models separately. We have ensemble of 8 models that gave us 0.414 on public, and 0.392 on private.", "1464139": "Thanks for sharing the solution and congratulations, mate.", "1464201": "Congrats!! seems your study level models ensemble were quite strong!! good job \n1. do you know the CV, LB scores of your OD part ? \n\n2. any external data used ?", "1464241": "Thank you!\n\n1. 5 folds UniverseNet101: 57.7 / 57 / 54.8 / 54 / 53 mAP (images without bboxes are kept). LB: 0.096 / 0.096 / 0.094 / 0.096 / 0.096.\n\n2. No external dataset.", "1464369": "congrats @aidynub 🎉🎉" }, "source": "meta" }