{ "id": 298030, "title": "22th place solution / for English edition and 中文版本", "url": "/competitions/sartorius-cell-instance-segmentation/writeups/never-ever-give-up-22th-place-solution-for-english", "author_name": "", "post_date": "2021-12-31T08:41:38.043Z", "votes": 20, "comment_count": 10, "views": 0, "content": "

Thanks to my teammates in this comp ! @wgz123 @deeeeeeeplearning @shajiayu They do a lot in this comp!

\n

English edition

\n

SOTA: Mask-RCNN
\nbackbone: SwinTransformer
\nanchor_size: mutil anchor size for different sortation
\nanchor_ratio: different ratio for different sortation
\npre-trained model: by using outer data provided by this comp
\ndata augmentations: Resize(), Pad(), Grid(), RandomFlip()
\nTTA: Resize(), Pad(), RandomFlip()
\nLr-plan: CosineAnnealing
\nFour models ensemble: One model for classifier and the rest for segmentation
\nswa: Stochastic Weights Averaging
\nscore_thr: min score for each instance
\nmin_pixel: min pixel for each instance
\nfix overlap: fix overlap by the index

\n

中文版本

\n

SOTA：Mask-RCNN
\n骨干网：SwinTransformer
\n框大小策略：不同类别采用不同的框大小
\n宽高比策略：统计标注种类宽高比分布
\n预训练：官方提供的外部数据进行预训练
\n数据增强：Resize(), Pad(), Grid(), RandomFlip()
\nTTA：Resize(), Pad(), RandomFlip()
\n学习率策略：余弦学习率
\n四模型集成：一模型分类三模型分割
\n随机加权平均：swa
\n卡置信度阈值：min_score
\n卡最小像素阈值：min_pixel
\n去除重叠：按照索引顺序去除

\n

Second comp in kaggle! keeping fighting!

", "messages": [ { "id": "1633985", "postDate": "12/31/2021 08:36:16", "content": "

Thanks to my teammates in this comp ! @wgz123 @deeeeeeeplearning @shajiayu They do a lot in this comp!

\n

English edition

\n

SOTA: Mask-RCNN
\nbackbone: SwinTransformer
\nanchor_size: mutil anchor size for different sortation
\nanchor_ratio: different ratio for different sortation
\npre-trained model: by using outer data provided by this comp
\ndata augmentations: Resize(), Pad(), Grid(), RandomFlip()
\nTTA: Resize(), Pad(), RandomFlip()
\nLr-plan: CosineAnnealing
\nFour models ensemble: One model for classifier and the rest for segmentation
\nswa: Stochastic Weights Averaging
\nscore_thr: min score for each instance
\nmin_pixel: min pixel for each instance
\nfix overlap: fix overlap by the index

\n

中文版本

\n

SOTA：Mask-RCNN
\n骨干网：SwinTransformer
\n框大小策略：不同类别采用不同的框大小
\n宽高比策略：统计标注种类宽高比分布
\n预训练：官方提供的外部数据进行预训练
\n数据增强：Resize(), Pad(), Grid(), RandomFlip()
\nTTA：Resize(), Pad(), RandomFlip()
\n学习率策略：余弦学习率
\n四模型集成：一模型分类三模型分割
\n随机加权平均：swa
\n卡置信度阈值：min_score
\n卡最小像素阈值：min_pixel
\n去除重叠：按照索引顺序去除

\n

Second comp in kaggle! keeping fighting!

", "rawMarkdown": "Thanks to my teammates in this comp ! @wgz123 @deeeeeeeplearning @shajiayu They do a lot in this comp!\n\n#English edition\n`SOTA`: Mask-RCNN\n`backbone`: SwinTransformer\n`anchor_size`: mutil anchor size for different sortation\n`anchor_ratio`: different ratio for different sortation\n`pre-trained model`: by using outer data provided by this comp\n`data augmentations`: Resize(), Pad(), Grid(), RandomFlip()\n`TTA`: Resize(), Pad(), RandomFlip()\n`Lr-plan`: CosineAnnealing\n`Four models ensemble`: One model for classifier and the rest for segmentation\n`swa`: Stochastic Weights Averaging\n`score_thr`: min score for each instance\n`min_pixel`: min pixel for each instance\n`fix overlap`: fix overlap by the index\n\n#中文版本\n`SOTA`：Mask-RCNN\n`骨干网`：SwinTransformer\n`框大小策略`：不同类别采用不同的框大小\n`宽高比策略`：统计标注种类宽高比分布\n`预训练`：官方提供的外部数据进行预训练\n`数据增强`：Resize(), Pad(), Grid(), RandomFlip()\n`TTA`：Resize(), Pad(), RandomFlip()\n`学习率策略`：余弦学习率 \n`四模型集成`：一模型分类三模型分割\n`随机加权平均`：swa\n`卡置信度阈值`：min_score\n`卡最小像素阈值`：min_pixel\n`去除重叠`：按照索引顺序去除\n\n\nSecond comp in kaggle! keeping fighting!", "votes": null }, { "id": "1633999", "postDate": "12/31/2021 08:46:07", "content": "

Good job! keep moving!

", "rawMarkdown": "Good job! keep moving!", "votes": null }, { "id": "1634265", "postDate": "12/31/2021 13:42:52", "content": "

Great work! What is the library that your team used? Is this MMdet?

", "rawMarkdown": "Great work! What is the library that your team used? Is this MMdet?", "votes": null }, { "id": "1634375", "postDate": "12/31/2021 16:44:21", "content": "

wow, you used SwinTransformer as a backbone, this is great. Are you guys planning to publish the code? Would love to check that out. Thank you.

", "rawMarkdown": "wow, you used `SwinTransformer` as a backbone, this is great. Are you guys planning to publish the code? Would love to check that out. Thank you.", "votes": null }, { "id": "1634517", "postDate": "12/31/2021 18:43:23", "content": "

Nice work, Manyu! May I know what is the size of your swin-t? And any specific model parameters you adjusted to use this backbone? I have try to adopt Swin-L in detectron, while it didn't work well because I am likely not to adjust some model parameters but using the default ones.

", "rawMarkdown": "Nice work, Manyu! May I know what is the size of your swin-t? And any specific model parameters you adjusted to use this backbone? I have try to adopt Swin-L in detectron, while it didn't work well because I am likely not to adjust some model parameters but using the default ones.", "votes": null }, { "id": "1634791", "postDate": "01/01/2022 07:06:20", "content": "

Right，mmdetection

", "rawMarkdown": "Right，mmdetection", "votes": null }, { "id": "1634792", "postDate": "01/01/2022 07:07:25", "content": "

I use the Resize() to resize the shorter edge from 800 to 1000

", "rawMarkdown": "I use the Resize() to resize the shorter edge from 800 to 1000", "votes": null }, { "id": "1634794", "postDate": "01/01/2022 07:08:17", "content": "

I will have a discussion with my teammates whether to publish our code

", "rawMarkdown": "I will have a discussion with my teammates whether to publish our code", "votes": null }, { "id": "1635457", "postDate": "01/01/2022 19:12:04", "content": "

Thank you for answering.

", "rawMarkdown": "Thank you for answering.", "votes": null }, { "id": "1736185", "postDate": "03/27/2022 04:48:43", "content": "

Can you make your code public?

", "rawMarkdown": "Can you make your code public?", "votes": null }, { "id": "1776721", "postDate": "05/04/2022 06:46:44", "content": "

so you use three seg models for three types of cell respectively instead of one model for three? I think Mask-RCNN is able to do seg for multi-class, did you do comparison experiments for the perf diff between them?

", "rawMarkdown": "so you use three seg models for three types of cell respectively instead of one model for three? I think Mask-RCNN is able to do seg for multi-class, did you do comparison experiments for the perf diff between them?", "votes": null } ], "comments": [ { "id": 1633999, "author_name": "guohey", "author_url": "", "post_date": "12/31/2021 08:46:07", "content": "

Good job! keep moving!

", "votes": null, "replies": [] }, { "id": 1634265, "author_name": "aengusng", "author_url": "", "post_date": "12/31/2021 13:42:52", "content": "

Great work! What is the library that your team used? Is this MMdet?

", "votes": null, "replies": [ { "id": 1634791, "author_name": "manyuli", "author_url": "", "post_date": "01/01/2022 07:06:20", "content": "

Right，mmdetection

", "votes": null, "replies": [] } ] }, { "id": 1634375, "author_name": "soumya9977", "author_url": "", "post_date": "12/31/2021 16:44:21", "content": "

wow, you used SwinTransformer as a backbone, this is great. Are you guys planning to publish the code? Would love to check that out. Thank you.

", "votes": null, "replies": [ { "id": 1634794, "author_name": "manyuli", "author_url": "", "post_date": "01/01/2022 07:08:17", "content": "

I will have a discussion with my teammates whether to publish our code

", "votes": null, "replies": [] }, { "id": 1635457, "author_name": "soumya9977", "author_url": "", "post_date": "01/01/2022 19:12:04", "content": "

Thank you for answering.

", "votes": null, "replies": [] } ] }, { "id": 1634517, "author_name": "charonwangg", "author_url": "", "post_date": "12/31/2021 18:43:23", "content": "

Nice work, Manyu! May I know what is the size of your swin-t? And any specific model parameters you adjusted to use this backbone? I have try to adopt Swin-L in detectron, while it didn't work well because I am likely not to adjust some model parameters but using the default ones.

", "votes": null, "replies": [ { "id": 1634792, "author_name": "manyuli", "author_url": "", "post_date": "01/01/2022 07:07:25", "content": "

I use the Resize() to resize the shorter edge from 800 to 1000

", "votes": null, "replies": [] } ] }, { "id": 1736185, "author_name": "zaopolearning", "author_url": "", "post_date": "03/27/2022 04:48:43", "content": "

Can you make your code public?

", "votes": null, "replies": [] }, { "id": 1776721, "author_name": "plugin1689", "author_url": "", "post_date": "05/04/2022 06:46:44", "content": "

so you use three seg models for three types of cell respectively instead of one model for three? I think Mask-RCNN is able to do seg for multi-class, did you do comparison experiments for the perf diff between them?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1633985": "Thanks to my teammates in this comp ! @wgz123 @deeeeeeeplearning @shajiayu They do a lot in this comp!\n\n#English edition\n`SOTA`: Mask-RCNN\n`backbone`: SwinTransformer\n`anchor_size`: mutil anchor size for different sortation\n`anchor_ratio`: different ratio for different sortation\n`pre-trained model`: by using outer data provided by this comp\n`data augmentations`: Resize(), Pad(), Grid(), RandomFlip()\n`TTA`: Resize(), Pad(), RandomFlip()\n`Lr-plan`: CosineAnnealing\n`Four models ensemble`: One model for classifier and the rest for segmentation\n`swa`: Stochastic Weights Averaging\n`score_thr`: min score for each instance\n`min_pixel`: min pixel for each instance\n`fix overlap`: fix overlap by the index\n\n#中文版本\n`SOTA`：Mask-RCNN\n`骨干网`：SwinTransformer\n`框大小策略`：不同类别采用不同的框大小\n`宽高比策略`：统计标注种类宽高比分布\n`预训练`：官方提供的外部数据进行预训练\n`数据增强`：Resize(), Pad(), Grid(), RandomFlip()\n`TTA`：Resize(), Pad(), RandomFlip()\n`学习率策略`：余弦学习率 \n`四模型集成`：一模型分类三模型分割\n`随机加权平均`：swa\n`卡置信度阈值`：min_score\n`卡最小像素阈值`：min_pixel\n`去除重叠`：按照索引顺序去除\n\n\nSecond comp in kaggle! keeping fighting!", "1633999": "Good job! keep moving!", "1634265": "Great work! What is the library that your team used? Is this MMdet?", "1634375": "wow, you used `SwinTransformer` as a backbone, this is great. Are you guys planning to publish the code? Would love to check that out. Thank you.", "1634517": "Nice work, Manyu! May I know what is the size of your swin-t? And any specific model parameters you adjusted to use this backbone? I have try to adopt Swin-L in detectron, while it didn't work well because I am likely not to adjust some model parameters but using the default ones.", "1634791": "Right，mmdetection", "1634792": "I use the Resize() to resize the shorter edge from 800 to 1000", "1634794": "I will have a discussion with my teammates whether to publish our code", "1635457": "Thank you for answering.", "1736185": "Can you make your code public?", "1776721": "so you use three seg models for three types of cell respectively instead of one model for three? I think Mask-RCNN is able to do seg for multi-class, did you do comparison experiments for the perf diff between them?" }, "source": "meta" }