Without knowing what the final result would be, I've decided to open-source one of my final submission kernels. It is a combination of two EfficientDet-D6 models with pseudo-labeling.
\nFirst model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and STAC procedure (\\lambda1 = 0.6, \\lambda2 = 0.4)
\nSecond model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and SplitBatchNorm from the AdvProp paper (splits = 2)
\nthe kernel is available here: https://www.kaggle.com/alexanderliao/effdet-d6-pl-s-bn-r-bb-a3-usa-eval-94-13-db
\nthe cleaned jigsaw dataset is here: https://www.kaggle.com/alexanderliao/wheatfullsize ; I fused all boxes over the edge by an automatic procedure. This brought about a significant change in model performance.
\nI also re-labeled the SPIKE dataset with a weaker model here because I thought the original quality was too poor: https://www.kaggle.com/alexanderliao/spike-pseudolabel . I did not succeed in incorporating them into my pipeline.
", "messages": [ { "id": "968747", "postDate": "08/13/2020 08:10:57", "content": "Without knowing what the final result would be, I've decided to open-source one of my final submission kernels. It is a combination of two EfficientDet-D6 models with pseudo-labeling.
\nFirst model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and STAC procedure (\\lambda1 = 0.6, \\lambda2 = 0.4)
\nSecond model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and SplitBatchNorm from the AdvProp paper (splits = 2)
\nthe kernel is available here: https://www.kaggle.com/alexanderliao/effdet-d6-pl-s-bn-r-bb-a3-usa-eval-94-13-db
\nthe cleaned jigsaw dataset is here: https://www.kaggle.com/alexanderliao/wheatfullsize ; I fused all boxes over the edge by an automatic procedure. This brought about a significant change in model performance.
\nI also re-labeled the SPIKE dataset with a weaker model here because I thought the original quality was too poor: https://www.kaggle.com/alexanderliao/spike-pseudolabel . I did not succeed in incorporating them into my pipeline.
", "rawMarkdown": "Without knowing what the final result would be, I've decided to open-source one of my final submission kernels. It is a combination of two EfficientDet-D6 models with pseudo-labeling.\n\nFirst model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and STAC procedure (\\lambda_1 = 0.6, \\lambda_2 = 0.4)\n\nSecond model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and SplitBatchNorm from the AdvProp paper (splits = 2)\n\nthe kernel is available here: https://www.kaggle.com/alexanderliao/effdet-d6-pl-s-bn-r-bb-a3-usa-eval-94-13-db\n\nthe cleaned jigsaw dataset is here: https://www.kaggle.com/alexanderliao/wheatfullsize ; I fused all boxes over the edge by an automatic procedure. This brought about a significant change in model performance.\n\nI also re-labeled the SPIKE dataset with a weaker model here because I thought the original quality was too poor: https://www.kaggle.com/alexanderliao/spike-pseudolabel . I did not succeed in incorporating them into my pipeline.", "votes": null }, { "id": "968751", "postDate": "08/13/2020 08:13:59", "content": "Thanks for sharing!
\n@alexanderliao
Awesome! You said that cleaning the dataset, i.e. patching together different parts, gave you a significant boost. Do you remember how much it was?
", "rawMarkdown": "Awesome! You said that cleaning the dataset, i.e. patching together different parts, gave you a significant boost. Do you remember how much it was?", "votes": null }, { "id": "969762", "postDate": "08/13/2020 23:05:11", "content": "better place to share will be github!
", "rawMarkdown": "better place to share will be github!", "votes": null }, { "id": "969939", "postDate": "08/14/2020 04:43:27", "content": "Informative. Thanks for sharing.
", "rawMarkdown": "Informative. Thanks for sharing.", "votes": null }, { "id": "970436", "postDate": "08/14/2020 13:02:00", "content": "Thanks and good luck with the final leaderboard. Whatever the final result you added a lot to this competition and certainly gave our team a lot of useful pointers.
", "rawMarkdown": "Thanks and good luck with the final leaderboard. Whatever the final result you added a lot to this competition and certainly gave our team a lot of useful pointers.", "votes": null }, { "id": "971464", "postDate": "08/15/2020 14:56:43", "content": "Thanks for sharing @alexanderliao Two questions.
\nI didn't take care about it because making my notebook deterministic gave consistently low behaviors. So I didn't bother to comment it out.
I forgot the time it took to submit, but it may need more than 6 hours to run. The commit time is less than 6 hours though.
thanks for replying so promptly
", "rawMarkdown": "thanks for replying so promptly", "votes": null }, { "id": "974161", "postDate": "08/17/2020 18:59:28", "content": "Thank you a lot, that's highly appreciated!
", "rawMarkdown": "Thank you a lot, that's highly appreciated!", "votes": null }, { "id": "974361", "postDate": "08/17/2020 23:44:08", "content": "Cool, thanks.
", "rawMarkdown": "Cool, thanks.", "votes": null }, { "id": "978081", "postDate": "08/19/2020 23:25:42", "content": "@alexanderliao
\nBeen an extremely tough competition, but congrats on your solo gold and GM!!
Thanks for sharing @alexanderliao and congrats GM on your solo gold
", "rawMarkdown": "Thanks for sharing @alexanderliao and congrats GM on your solo gold", "votes": null } ], "comments": [ { "id": 968751, "author_name": "piantic", "author_url": "", "post_date": "08/13/2020 08:13:59", "content": "Thanks for sharing!
\n@alexanderliao
Awesome! You said that cleaning the dataset, i.e. patching together different parts, gave you a significant boost. Do you remember how much it was?
", "votes": null, "replies": [] }, { "id": 969762, "author_name": "jiwongpark", "author_url": "", "post_date": "08/13/2020 23:05:11", "content": "better place to share will be github!
", "votes": null, "replies": [] }, { "id": 970436, "author_name": "datahobbit", "author_url": "", "post_date": "08/14/2020 13:02:00", "content": "Thanks and good luck with the final leaderboard. Whatever the final result you added a lot to this competition and certainly gave our team a lot of useful pointers.
", "votes": null, "replies": [] }, { "id": 971464, "author_name": "nvnnghia", "author_url": "", "post_date": "08/15/2020 14:56:43", "content": "Thanks for sharing @alexanderliao Two questions.
\nI didn't take care about it because making my notebook deterministic gave consistently low behaviors. So I didn't bother to comment it out.
I forgot the time it took to submit, but it may need more than 6 hours to run. The commit time is less than 6 hours though.
thanks for replying so promptly
", "votes": null, "replies": [] } ] }, { "id": 974361, "author_name": "agarwalarti", "author_url": "", "post_date": "08/17/2020 23:44:08", "content": "Cool, thanks.
", "votes": null, "replies": [] }, { "id": 978081, "author_name": "kyoshioka47", "author_url": "", "post_date": "08/19/2020 23:25:42", "content": "@alexanderliao
\nBeen an extremely tough competition, but congrats on your solo gold and GM!!
Thanks for sharing @alexanderliao and congrats GM on your solo gold
", "votes": null, "replies": [] }, { "id": 969939, "author_name": "raoofnaushad", "author_url": "", "post_date": "08/14/2020 04:43:27", "content": "Informative. Thanks for sharing.
", "votes": null, "replies": [] }, { "id": 974161, "author_name": "ChristianDenich", "author_url": "", "post_date": "08/17/2020 18:59:28", "content": "Thank you a lot, that's highly appreciated!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "968747": "Without knowing what the final result would be, I've decided to open-source one of my final submission kernels. It is a combination of two EfficientDet-D6 models with pseudo-labeling.\n\nFirst model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and STAC procedure (\\lambda_1 = 0.6, \\lambda_2 = 0.4)\n\nSecond model: EfficientDet-D6, pretrained with coco, trained 1st round with original data, second round with cleaned jigsaw data and SplitBatchNorm from the AdvProp paper (splits = 2)\n\nthe kernel is available here: https://www.kaggle.com/alexanderliao/effdet-d6-pl-s-bn-r-bb-a3-usa-eval-94-13-db\n\nthe cleaned jigsaw dataset is here: https://www.kaggle.com/alexanderliao/wheatfullsize ; I fused all boxes over the edge by an automatic procedure. This brought about a significant change in model performance.\n\nI also re-labeled the SPIKE dataset with a weaker model here because I thought the original quality was too poor: https://www.kaggle.com/alexanderliao/spike-pseudolabel . I did not succeed in incorporating them into my pipeline.", "968751": "Thanks for sharing!\n@alexanderliao", "969558": "Awesome! You said that cleaning the dataset, i.e. patching together different parts, gave you a significant boost. Do you remember how much it was?", "969762": "better place to share will be github!", "969939": "Informative. Thanks for sharing.", "970436": "Thanks and good luck with the final leaderboard. Whatever the final result you added a lot to this competition and certainly gave our team a lot of useful pointers.", "971464": "Thanks for sharing @alexanderliao Two questions. \n1. In [this thread](https://www.kaggle.com/c/global-wheat-detection/discussion/172229) you said that setting both cudnn's benchmark and determinism to True is technically conflicting, However, both cudnn's benchmark and determinism is set to True in your notebook. so, which is the correct way to set seed?\n2. How long does it take to submit your notebook? I think it needs more than 6 hours.", "971467": "1. I didn't take care about it because making my notebook deterministic gave consistently low behaviors. So I didn't bother to comment it out.\n\n2. I forgot the time it took to submit, but it may need more than 6 hours to run. The commit time is less than 6 hours though.", "971472": "thanks for replying so promptly", "974161": "Thank you a lot, that's highly appreciated!", "974361": "Cool, thanks.", "978081": "alexanderliao \nBeen an extremely tough competition, but congrats on your solo gold and GM!!", "978171": "Thanks for sharing @alexanderliao and congrats GM on your solo gold" }, "source": "meta" }