{ "id": 428301, "title": "10th place solutions!", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/428301", "author_name": "suguuuuu", "post_date": "2023-08-01T00:56:14.698000", "votes": 18, "comment_count": 8, "views": 0, "content": "

Thank you for host and competitors.
\nIt is interesting competition for me.

\n

I'm happy to become a Kaggle master!!

\n

Team Member

\n

@sugupoko, @hatakee, @kfuji
\nco-workers!!

\n

Solution overview

\n

\"\"

\n

Thank you @fnands !!
\nbase notebook : https://www.kaggle.com/code/fnands/a-quick-yolov7-baseline

\n

Yolo's segmentation head size is 1/4 (512 => 128). So I changed the input resolution( 640 => 160).

\n

Unique idea (maybe)

\n

I did fine-tuning with ds1 in the last, because I didn't want to use dilate !!

\n

Inference code

\n

https://www.kaggle.com/sugupoko/10th-place-inference-yolov7ensemble-pseudo-res640

\n

Reference previous competition

\n

Our many experiments are from the previous two competitions

\n\n

Things that didn't go well:

\n\n

Things I couldn't implement well:

\n\n

Things I couldn't do:

\n\n

Other

\n

I trained higher resolution(640=>800). That LB is the highest of our models.
\nLB: 0.473, PB0.563

", "messages": [ { "id": 2367989, "postDate": "2023-08-01T00:56:14.700Z", "content": "

Thank you for host and competitors.
\nIt is interesting competition for me.

\n

I'm happy to become a Kaggle master!!

\n

Team Member

\n

@sugupoko, @hatakee, @kfuji
\nco-workers!!

\n

Solution overview

\n

\"\"

\n

Thank you @fnands !!
\nbase notebook : https://www.kaggle.com/code/fnands/a-quick-yolov7-baseline

\n

Yolo's segmentation head size is 1/4 (512 => 128). So I changed the input resolution( 640 => 160).

\n

Unique idea (maybe)

\n

I did fine-tuning with ds1 in the last, because I didn't want to use dilate !!

\n

Inference code

\n

https://www.kaggle.com/sugupoko/10th-place-inference-yolov7ensemble-pseudo-res640

\n

Reference previous competition

\n

Our many experiments are from the previous two competitions

\n\n

Things that didn't go well:

\n\n

Things I couldn't implement well:

\n\n

Things I couldn't do:

\n\n

Other

\n

I trained higher resolution(640=>800). That LB is the highest of our models.
\nLB: 0.473, PB0.563

", "rawMarkdown": "Thank you for host and competitors.\nIt is interesting competition for me.\n\nI'm happy to become a Kaggle master!!\n\n## Team Member\n@sugupoko, @hatakee, @kfuji\nco-workers!!\n\n## Solution overview\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fec3da742a7a4a7c70bd8d56634204885%2Fsol.png?generation=1690850953870523&alt=media)\n\nThank you @fnands !!\nbase notebook : https://www.kaggle.com/code/fnands/a-quick-yolov7-baseline\n\nYolo's segmentation head size is 1/4 (512 => 128). So I changed the input resolution( 640 => 160).\n\n### Unique idea (maybe)\nI did fine-tuning with ds1 in the last, because **I didn't want to use dilate !!**\n\n## Inference code\nhttps://www.kaggle.com/sugupoko/10th-place-inference-yolov7ensemble-pseudo-res640\n\n## Reference previous competition\nOur many experiments are from the previous two competitions\n- https://www.kaggle.com/competitions/global-wheat-detection\n- https://www.kaggle.com/c/sartorius-cell-instance-segmentation/overview\n\n## Things that didn't go well:\n- Applied pseudo-labeling twice.\n- Yolov8 (training and inference).\n- mmdetection. (difficult for us...)\n- Unet prediction in the detected BBOX.\n- NMS => WSF \n - https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion\n\n## Things I couldn't implement well:\n- Training with data regenerated by combining tiles.\n\n## Things I couldn't do:\n- Augmentation with the stain tool.\n- Learning using extra data.\n\n## Other\nI trained higher resolution(640=>800). That LB is the highest of our models.\nLB: 0.473, PB0.563", "votes": 16 }, { "id": 2368344, "postDate": "2023-08-01T06:04:42.833Z", "content": "

Very interesting pipelines. Would you develop this into a complete writeup and compete for the 100K prize?

", "rawMarkdown": "Very interesting pipelines. Would you develop this into a complete writeup and compete for the 100K prize?", "votes": 1, "replies": [ { "id": 2368397, "postDate": "2023-08-01T06:35:23.537Z", "content": "

Thank you for your suggestion. I don't know that prize! I'll try!

", "rawMarkdown": "Thank you for your suggestion. I don't know that prize! I'll try!", "votes": 1, "replies": [ { "id": 2369042, "postDate": "2023-08-01T14:00:14.880Z", "content": "

Here is the anouncement: https://www.kaggle.com/discussions/general/427114?utm_medium=email&utm_source=gamma&utm_campaign=solutionwriteup-2023

", "rawMarkdown": "Here is the anouncement: https://www.kaggle.com/discussions/general/427114?utm_medium=email&utm_source=gamma&utm_campaign=solutionwriteup-2023", "votes": 1 } ] } ] }, { "id": 2367997, "postDate": "2023-08-01T01:07:48.690Z", "content": "

I did fine-tuning with ds1 because I didn't want to use dilate.

", "rawMarkdown": "I did fine-tuning with ds1 because I didn't want to use dilate.", "votes": 1 }, { "id": 2452813, "postDate": "2023-09-23T15:44:01.310Z", "content": "

Oh cool! Congrats!

", "rawMarkdown": "Oh cool! Congrats!" }, { "id": 2371874, "postDate": "2023-08-03T10:12:39.850Z", "content": "

May I know the impact of your post processing on the LB/private score

", "rawMarkdown": "May I know the impact of your post processing on the LB/private score", "replies": [ { "id": 2373568, "postDate": "2023-08-04T10:04:41.597Z", "content": "

The effect was quite minimal, about +0.001 …

", "rawMarkdown": "The effect was quite minimal, about +0.001 ..." } ] }, { "id": 2368048, "postDate": "2023-08-01T01:52:24.353Z", "content": "

Congratulations! Thanks for sharing your solid work!
\nPS: Diation is a vvvvvery failed method,
\nand stain tool is useful for augmentation.

", "rawMarkdown": "Congratulations! Thanks for sharing your solid work!\nPS: Diation is a vvvvvery failed method,\nand stain tool is useful for augmentation.", "votes": 1, "isDeleted": true } ], "comments": [ { "id": 2368344, "author_name": "Gabriel Preda", "author_url": "", "post_date": "2023-08-01T06:04:42.833000", "content": "

Very interesting pipelines. Would you develop this into a complete writeup and compete for the 100K prize?

", "votes": 1, "replies": [ { "id": 2368397, "author_name": "suguuuuu", "author_url": "", "post_date": "2023-08-01T06:35:23.537000", "content": "

Thank you for your suggestion. I don't know that prize! I'll try!

", "votes": 1, "replies": [ { "id": 2369042, "author_name": "Gabriel Preda", "author_url": "", "post_date": "2023-08-01T14:00:14.880000", "content": "

Here is the anouncement: https://www.kaggle.com/discussions/general/427114?utm_medium=email&utm_source=gamma&utm_campaign=solutionwriteup-2023

", "votes": 1, "replies": [] } ] } ] }, { "id": 2367997, "author_name": "suguuuuu", "author_url": "", "post_date": "2023-08-01T01:07:48.690000", "content": "

I did fine-tuning with ds1 because I didn't want to use dilate.

", "votes": 1, "replies": [] }, { "id": 2452813, "author_name": "fnands", "author_url": "", "post_date": "2023-09-23T15:44:01.310000", "content": "

Oh cool! Congrats!

", "votes": 0, "replies": [] }, { "id": 2371874, "author_name": "Dive Deeper", "author_url": "", "post_date": "2023-08-03T10:12:39.850000", "content": "

May I know the impact of your post processing on the LB/private score

", "votes": 0, "replies": [ { "id": 2373568, "author_name": "suguuuuu", "author_url": "", "post_date": "2023-08-04T10:04:41.597000", "content": "

The effect was quite minimal, about +0.001 …

", "votes": 0, "replies": [] } ] }, { "id": 2368048, "author_name": "", "author_url": "", "post_date": "2023-08-01T01:52:24.353000", "content": "

Congratulations! Thanks for sharing your solid work!
\nPS: Diation is a vvvvvery failed method,
\nand stain tool is useful for augmentation.

", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "2367989": "Thank you for host and competitors.\nIt is interesting competition for me.\n\nI'm happy to become a Kaggle master!!\n\n## Team Member\n@sugupoko, @hatakee, @kfuji\nco-workers!!\n\n## Solution overview\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2Fec3da742a7a4a7c70bd8d56634204885%2Fsol.png?generation=1690850953870523&alt=media)\n\nThank you @fnands !!\nbase notebook : https://www.kaggle.com/code/fnands/a-quick-yolov7-baseline\n\nYolo's segmentation head size is 1/4 (512 => 128). So I changed the input resolution( 640 => 160).\n\n### Unique idea (maybe)\nI did fine-tuning with ds1 in the last, because **I didn't want to use dilate !!**\n\n## Inference code\nhttps://www.kaggle.com/sugupoko/10th-place-inference-yolov7ensemble-pseudo-res640\n\n## Reference previous competition\nOur many experiments are from the previous two competitions\n- https://www.kaggle.com/competitions/global-wheat-detection\n- https://www.kaggle.com/c/sartorius-cell-instance-segmentation/overview\n\n## Things that didn't go well:\n- Applied pseudo-labeling twice.\n- Yolov8 (training and inference).\n- mmdetection. (difficult for us...)\n- Unet prediction in the detected BBOX.\n- NMS => WSF \n - https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion\n\n## Things I couldn't implement well:\n- Training with data regenerated by combining tiles.\n\n## Things I couldn't do:\n- Augmentation with the stain tool.\n- Learning using extra data.\n\n## Other\nI trained higher resolution(640=>800). That LB is the highest of our models.\nLB: 0.473, PB0.563", "2368344": "Very interesting pipelines. Would you develop this into a complete writeup and compete for the 100K prize?", "2367997": "I did fine-tuning with ds1 because I didn't want to use dilate.", "2452813": "Oh cool! Congrats!", "2371874": "May I know the impact of your post processing on the LB/private score", "2368048": "Congratulations! Thanks for sharing your solid work!\nPS: Diation is a vvvvvery failed method,\nand stain tool is useful for augmentation." } }