{ "id": 420063, "title": "LB : 0.377(single model) and 0.382(ensemble) with YoLoV7", "url": "/competitions/hubmap-hacking-the-human-vasculature/discussion/420063", "author_name": "", "post_date": "2023-06-29T04:27:40.436194700Z", "votes": 4, "comment_count": 11, "views": 0, "content": "

Current Setting based on fnands's notebooks and thank you very much

\n

Things works for me:

\n
    \n
  1. training model without \"unsure\"
    2. \n
\n

ToDo:

\n
    \n
  1. use model to predict the unsure lable area
    2. \n
  2. Resolution: different image size
    3. \n
  3. Detectron2 model
    4. \n
  4. merge resutls with other models such as detectron2, transformer or Vit-based models
    5. \n
  5. try better augmentation combinations
    6. \n
  6. [find a better data split and multi-fold evaluation]
    \n(https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419980)
    7. \n
  7. tune some thresholds and lr, loss
    8. \n
  8. use pretrained model with only one or two classes to turn the model again
    9. \n
  9. TTA
    10. \n
  10. Dilate
    11. \n
\n
\n

learned from hengck23, I decide to open a discussion to record my experiment results

\n
", "messages": [ { "id": "2322079", "postDate": "06/29/2023 04:27:40", "content": "

Current Setting based on fnands's notebooks and thank you very much

\n

Things works for me:

\n
    \n
  1. training model without \"unsure\"
    2. \n
\n

ToDo:

\n
    \n
  1. use model to predict the unsure lable area
    2. \n
  2. Resolution: different image size
    3. \n
  3. Detectron2 model
    4. \n
  4. merge resutls with other models such as detectron2, transformer or Vit-based models
    5. \n
  5. try better augmentation combinations
    6. \n
  6. [find a better data split and multi-fold evaluation]
    \n(https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419980)
    7. \n
  7. tune some thresholds and lr, loss
    8. \n
  8. use pretrained model with only one or two classes to turn the model again
    9. \n
  9. TTA
    10. \n
  10. Dilate
    11. \n
\n
\n

learned from hengck23, I decide to open a discussion to record my experiment results

\n
", "rawMarkdown": "# Current Setting based on [fnands's ](https://www.kaggle.com/fnands) notebooks and thank you very much\n\n\n# Things works for me:\n1. training model without \"unsure\"\n\n\n\n\n**ToDo**:\n1. use model to predict the unsure lable area\n2. Resolution: different image size\n3. Detectron2 model\n4. [merge resutls with other models such as detectron2, transformer or Vit-based models](https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419857)\n5. try better augmentation combinations\n6. [find a better data split and multi-fold evaluation]\n(https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419980)\n7. tune[ some thresholds](https://www.kaggle.com/code/allenwpr/hubmap-yolov7-ensemble) and lr, loss\n8. use pretrained model with only one or two classes to turn the model again\n9. TTA\n10. Dilate\n\n\n>learned from hengck23, I decide to open a discussion to record my experiment results", "votes": null }, { "id": "2322164", "postDate": "06/29/2023 05:51:51", "content": "

good job，I'm looking forward to the results of your experiment.

", "rawMarkdown": "good job，I'm looking forward to the results of your experiment.", "votes": null }, { "id": "2322713", "postDate": "06/29/2023 12:55:18", "content": "

👀👌 stay tune~

", "rawMarkdown": "👀👌 stay tune~", "votes": null }, { "id": "2322822", "postDate": "06/29/2023 14:10:53", "content": "

Hi. Are you using \"dilation\" of mask ?

\n

https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901

", "rawMarkdown": "Hi. Are you using \"dilation\" of mask ?\n\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901", "votes": null }, { "id": "2323399", "postDate": "06/29/2023 23:24:08", "content": "

Thank you for your recommendation.
\nI didn't use it yet,
\nBut I think I should try it.

\n

By the way, for inference, should we do the same thing?

", "rawMarkdown": "Thank you for your recommendation. \nI didn't use it yet, \nBut I think I should try it. \n\nBy the way, for inference, should we do the same thing?", "votes": null }, { "id": "2323412", "postDate": "06/29/2023 23:44:18", "content": "

Please refer to the “cv2.dilate()” part here. It is better to optimize the kernel size.
\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7

\n

Also, please refer to here for the validity of the method.
\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901

", "rawMarkdown": "Please refer to the “cv2.dilate()” part here. It is better to optimize the kernel size.\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7\n\nAlso, please refer to here for the validity of the method.\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901", "votes": null }, { "id": "2341252", "postDate": "07/12/2023 01:37:55", "content": "

Hey, your to-do list inspires me a lot. Have you tried use pretrained model with only one or two classes? I think it should be useful and I wonder how to find this kind of pretrained model.

", "rawMarkdown": "Hey, your to-do list inspires me a lot. Have you tried use pretrained model with only one or two classes? I think it should be useful and I wonder how to find this kind of pretrained model.", "votes": null }, { "id": "2341295", "postDate": "07/12/2023 03:01:10", "content": "

LB : 0.494(single model) with YoLoV7
\nHow did you make the model ensemble?I use 5 fold models for ensemble, but it's worse than a single model.
\nTTA doesn't work for me either

", "rawMarkdown": "LB : 0.494(single model) with YoLoV7\nHow did you make the model ensemble?I use 5 fold models for ensemble, but it's worse than a single model.\nTTA doesn't work for me either", "votes": null }, { "id": "2341850", "postDate": "07/12/2023 10:27:17", "content": "

You can refer to this notebook：https://www.kaggle.com/code/bachngoh/hubmap-yolov7-ensemble

", "rawMarkdown": "You can refer to this notebook：https://www.kaggle.com/code/bachngoh/hubmap-yolov7-ensemble", "votes": null }, { "id": "2349169", "postDate": "07/18/2023 08:02:14", "content": "

maybe this works:
\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"
\nWBF: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
\nWSF: https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook
\nrefer to https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419143#2318422 by hengck23

", "rawMarkdown": "maybe this works:\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"\nWBF: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview\nWSF: https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook\nrefer to https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419143#2318422 by hengck23", "votes": null }, { "id": "2350063", "postDate": "07/19/2023 00:29:11", "content": "

Thanks, i will try it.

", "rawMarkdown": "Thanks, i will try it.", "votes": null }, { "id": "2351600", "postDate": "07/20/2023 09:01:02", "content": "

both not work

", "rawMarkdown": "both not work", "votes": null } ], "comments": [ { "id": 2322164, "author_name": "qianxin1098", "author_url": "", "post_date": "06/29/2023 05:51:51", "content": "

good job，I'm looking forward to the results of your experiment.

", "votes": null, "replies": [ { "id": 2322713, "author_name": "chg0901", "author_url": "", "post_date": "06/29/2023 12:55:18", "content": "

👀👌 stay tune~

", "votes": null, "replies": [] } ] }, { "id": 2322822, "author_name": "mitsuyasuhoshino", "author_url": "", "post_date": "06/29/2023 14:10:53", "content": "

Hi. Are you using \"dilation\" of mask ?

\n

https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901

", "votes": null, "replies": [ { "id": 2323399, "author_name": "chg0901", "author_url": "", "post_date": "06/29/2023 23:24:08", "content": "

Thank you for your recommendation.
\nI didn't use it yet,
\nBut I think I should try it.

\n

By the way, for inference, should we do the same thing?

", "votes": null, "replies": [ { "id": 2323412, "author_name": "mitsuyasuhoshino", "author_url": "", "post_date": "06/29/2023 23:44:18", "content": "

Please refer to the “cv2.dilate()” part here. It is better to optimize the kernel size.
\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7

\n

Also, please refer to here for the validity of the method.
\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901

", "votes": null, "replies": [] } ] } ] }, { "id": 2341252, "author_name": "bluehe", "author_url": "", "post_date": "07/12/2023 01:37:55", "content": "

Hey, your to-do list inspires me a lot. Have you tried use pretrained model with only one or two classes? I think it should be useful and I wonder how to find this kind of pretrained model.

", "votes": null, "replies": [] }, { "id": 2341295, "author_name": "bent1e", "author_url": "", "post_date": "07/12/2023 03:01:10", "content": "

LB : 0.494(single model) with YoLoV7
\nHow did you make the model ensemble?I use 5 fold models for ensemble, but it's worse than a single model.
\nTTA doesn't work for me either

", "votes": null, "replies": [ { "id": 2341850, "author_name": "peilwang", "author_url": "", "post_date": "07/12/2023 10:27:17", "content": "

You can refer to this notebook：https://www.kaggle.com/code/bachngoh/hubmap-yolov7-ensemble

", "votes": null, "replies": [] }, { "id": 2349169, "author_name": "aldebaranhd", "author_url": "", "post_date": "07/18/2023 08:02:14", "content": "

maybe this works:
\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"
\nWBF: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview
\nWSF: https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook
\nrefer to https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419143#2318422 by hengck23

", "votes": null, "replies": [ { "id": 2351600, "author_name": "bent1e", "author_url": "", "post_date": "07/20/2023 09:01:02", "content": "

both not work

", "votes": null, "replies": [] } ] } ] }, { "id": 2350063, "author_name": "cruelkanade", "author_url": "", "post_date": "07/19/2023 00:29:11", "content": "

Thanks, i will try it.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2322079": "# Current Setting based on [fnands's ](https://www.kaggle.com/fnands) notebooks and thank you very much\n\n\n# Things works for me:\n1. training model without \"unsure\"\n\n\n\n\n**ToDo**:\n1. use model to predict the unsure lable area\n2. Resolution: different image size\n3. Detectron2 model\n4. [merge resutls with other models such as detectron2, transformer or Vit-based models](https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419857)\n5. try better augmentation combinations\n6. [find a better data split and multi-fold evaluation]\n(https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419980)\n7. tune[ some thresholds](https://www.kaggle.com/code/allenwpr/hubmap-yolov7-ensemble) and lr, loss\n8. use pretrained model with only one or two classes to turn the model again\n9. TTA\n10. Dilate\n\n\n>learned from hengck23, I decide to open a discussion to record my experiment results", "2322164": "good job，I'm looking forward to the results of your experiment.", "2322713": "👀👌 stay tune~", "2322822": "Hi. Are you using \"dilation\" of mask ?\n\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901", "2323399": "Thank you for your recommendation. \nI didn't use it yet, \nBut I think I should try it. \n\nBy the way, for inference, should we do the same thing?", "2323412": "Please refer to the “cv2.dilate()” part here. It is better to optimize the kernel size.\nhttps://www.kaggle.com/code/hengck23/lb4-09-baseline-yolov7\n\nAlso, please refer to here for the validity of the method.\nhttps://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/416901", "2341252": "Hey, your to-do list inspires me a lot. Have you tried use pretrained model with only one or two classes? I think it should be useful and I wonder how to find this kind of pretrained model.", "2341295": "LB : 0.494(single model) with YoLoV7\nHow did you make the model ensemble?I use 5 fold models for ensemble, but it's worse than a single model.\nTTA doesn't work for me either", "2341850": "You can refer to this notebook：https://www.kaggle.com/code/bachngoh/hubmap-yolov7-ensemble", "2349169": "maybe this works:\n\"Ensembling masks from different models with customed Weighted Boxes Fusion\"\nWBF: https://www.kaggle.com/competitions/sartorius-cell-instance-segmentation/overview\nWSF: https://www.kaggle.com/code/mistag/sartorius-tta-with-weighted-segments-fusion/notebook\nrefer to https://www.kaggle.com/competitions/hubmap-hacking-the-human-vasculature/discussion/419143#2318422 by hengck23", "2350063": "Thanks, i will try it.", "2351600": "both not work" }, "source": "meta" }