Hi everyone!
\n\nI would like to share with you new kernel about optimizing WBF params for EfficientDet!
\n\nWhy it may be useful for you? It gives really good boost without additional training.
\n\nI have got new scores:\n- 0.7176 --> 0.7193 single\n- 0.7294 --> 0.7330 cutmix --> new\n- 0.7318 --> 0.7338 cutmix + mixup\n- 0.7375 (ensemble of two above)
\n\nYou should change values of this params in your inference:
\n\n```
\n\nWrite in comment have got boost or not ;)
\n\nP.S. Also I would like to ask you, does it help anyone? Or everyone knows and uses any other values of this parameters? Do you want to know about improvements from me or I need stop with publishing? Thank you for answer in advance!
\n\n[EDITED]:
\n\nI have updated new version of kernel with optimization parameters for 4 methods: nms, soft_nms, nmw, wbf.
", "messages": [ { "id": "864688", "postDate": "05/28/2020 06:17:27", "content": "Hi everyone!
\n\nI would like to share with you new kernel about optimizing WBF params for EfficientDet!
\n\nWhy it may be useful for you? It gives really good boost without additional training.
\n\nI have got new scores:\n- 0.7176 --> 0.7193 single\n- 0.7294 --> 0.7330 cutmix --> new\n- 0.7318 --> 0.7338 cutmix + mixup\n- 0.7375 (ensemble of two above)
\n\nYou should change values of this params in your inference:
\n\n```
\n\nWrite in comment have got boost or not ;)
\n\nP.S. Also I would like to ask you, does it help anyone? Or everyone knows and uses any other values of this parameters? Do you want to know about improvements from me or I need stop with publishing? Thank you for answer in advance!
\n\n[EDITED]:
\n\nI have updated new version of kernel with optimization parameters for 4 methods: nms, soft_nms, nmw, wbf.
", "rawMarkdown": "Hi everyone!\n\nI would like to share with you [new kernel about optimizing WBF params for EfficientDet](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet)! \n\nWhy it may be useful for you? It gives really good boost without additional training.\n\nI have got new scores:\n- 0.7176 --> 0.7193 [single](https://www.kaggle.com/shonenkov/inference-efficientdet)\n- 0.7294 --> 0.7330 [cutmix](https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet?scriptVersionId=34505469) --> [new](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet?scriptVersionId=34928761)\n- 0.7318 --> 0.7338 [cutmix + mixup](https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet?scriptVersionId=34656263)\n- 0.7375 (ensemble of two above) \n\nYou should change values of this params in your inference:\n\n```\n------------------------------\n[Best Iou Thr]: 0.44\n[Best Skip Box Thr]: 0.43\n[Best Score]: 0.6953\n------------------------------\n```\n\nWrite in comment have got boost or not ;)\n\nP.S. Also I would like to ask you, does it help anyone? Or everyone knows and uses any other values of this parameters? Do you want to know about improvements from me or I need stop with publishing? Thank you for answer in advance! \n\n[EDITED]:\n\nI have updated new version of kernel with optimization parameters for 4 methods: nms, soft_nms, nmw, wbf.", "votes": null }, { "id": "864845", "postDate": "05/28/2020 08:17:36", "content": "I tried your train code, cant get that score...\nthanks for sharing, you are really imporving this compete.
", "rawMarkdown": "I tried your train code, cant get that score...\nthanks for sharing, you are really imporving this compete.", "votes": null }, { "id": "867435", "postDate": "05/30/2020 09:30:51", "content": "Great Kernel!\nLearning a lot from this competition. \nThanks:)
", "rawMarkdown": "Great Kernel!\nLearning a lot from this competition. \nThanks:)", "votes": null }, { "id": "867444", "postDate": "05/30/2020 09:48:14", "content": "welcome! 👍
", "rawMarkdown": "welcome! 👍", "votes": null }, { "id": "869844", "postDate": "06/01/2020 10:12:40", "content": "Very helpful kernel.
\n\nI wonder whether you train from scratch or fine-tune on each fold. I achieved a small boost but maybe i am missing the actual thing.
", "rawMarkdown": "Very helpful kernel.\n\nI wonder whether you train from scratch or fine-tune on each fold. I achieved a small boost but maybe i am missing the actual thing.", "votes": null }, { "id": "869869", "postDate": "06/01/2020 10:35:23", "content": "I don't know why this doesn't work for me\nBefore Bayesian Optimization : 0.7355\nAfter Bayesian Optimization : 0.7346
", "rawMarkdown": "I don't know why this doesn't work for me\nBefore Bayesian Optimization : 0.7355\nAfter Bayesian Optimization : 0.7346", "votes": null }, { "id": "869933", "postDate": "06/01/2020 11:39:46", "content": "@gyanaluckydas what models did you use? effdet?
", "rawMarkdown": "gyanaluckydas what models did you use? effdet?", "votes": null }, { "id": "869965", "postDate": "06/01/2020 12:05:31", "content": "Yes..effdet with pseudo label
", "rawMarkdown": "Yes..effdet with pseudo label", "votes": null }, { "id": "869974", "postDate": "06/01/2020 12:13:50", "content": "@gyanaluckydas pseudo labeling approach can shift confidences of models, for this case should run optimizing again
", "rawMarkdown": "gyanaluckydas pseudo labeling approach can shift confidences of models, for this case should run optimizing again", "votes": null }, { "id": "869998", "postDate": "06/01/2020 12:22:28", "content": "I can't understand...what do u mean by optimizing again
", "rawMarkdown": "I can't understand...what do u mean by optimizing again", "votes": null }, { "id": "870001", "postDate": "06/01/2020 12:23:41", "content": "bayesian-optimization-wbf-efficientdet with using your pseudo labeling. you should use val fold for pseudo-labeling and holdout fold for bayesian.
\n\nWelcome!
", "rawMarkdown": "[bayesian-optimization-wbf-efficientdet](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet) with using your pseudo labeling. you should use val fold for pseudo-labeling and holdout fold for bayesian.\n\nWelcome!", "votes": null }, { "id": "870024", "postDate": "06/01/2020 12:47:08", "content": "Thanks.. I will try it
", "rawMarkdown": "Thanks.. I will try it", "votes": null }, { "id": "875848", "postDate": "06/06/2020 07:52:18", "content": "Thanks, your kernels are very instructive!
", "rawMarkdown": "Thanks, your kernels are very instructive!", "votes": null }, { "id": "876211", "postDate": "06/06/2020 14:31:59", "content": "I've noticed that changing the iou_thr for soft_nms does not affect my accuracy at all, any idea why that could be?
Edit: Nevermind, after looking into the source code I understand now. You could remove that parameter from your soft_nms search
", "rawMarkdown": "I've noticed that changing the `iou_thr` for `soft_nms` does not affect my accuracy at all, any idea why that could be?\n\nEdit: Nevermind, after looking into the source code I understand now. You could remove that parameter from your soft_nms search", "votes": null }, { "id": "885459", "postDate": "06/14/2020 08:25:16", "content": "Hi @shonenkov really thank you for your kindly sharing. one question, all your mentioned scores are get with single model? such as 0.7294 --> 0.7330 cutmix
", "rawMarkdown": "Hi @shonenkov really thank you for your kindly sharing. one question, all your mentioned scores are get with single model? such as 0.7294 --> 0.7330 cutmix", "votes": null }, { "id": "885470", "postDate": "06/14/2020 08:43:24", "content": "@cswwp347724 thank you! this improvement for kfold
", "rawMarkdown": "cswwp347724 thank you! this improvement for kfold", "votes": null } ], "comments": [ { "id": 864845, "author_name": "tiandaye", "author_url": "", "post_date": "05/28/2020 08:17:36", "content": "I tried your train code, cant get that score...\nthanks for sharing, you are really imporving this compete.
", "votes": null, "replies": [] }, { "id": 867435, "author_name": "", "author_url": "", "post_date": "05/30/2020 09:30:51", "content": "Great Kernel!\nLearning a lot from this competition. \nThanks:)
", "votes": null, "replies": [ { "id": 867444, "author_name": "shonenkov", "author_url": "", "post_date": "05/30/2020 09:48:14", "content": "welcome! 👍
", "votes": null, "replies": [] } ] }, { "id": 869844, "author_name": "batuhangencay", "author_url": "", "post_date": "06/01/2020 10:12:40", "content": "Very helpful kernel.
\n\nI wonder whether you train from scratch or fine-tune on each fold. I achieved a small boost but maybe i am missing the actual thing.
", "votes": null, "replies": [] }, { "id": 869869, "author_name": "gyanaluckydas", "author_url": "", "post_date": "06/01/2020 10:35:23", "content": "I don't know why this doesn't work for me\nBefore Bayesian Optimization : 0.7355\nAfter Bayesian Optimization : 0.7346
", "votes": null, "replies": [ { "id": 869933, "author_name": "shonenkov", "author_url": "", "post_date": "06/01/2020 11:39:46", "content": "@gyanaluckydas what models did you use? effdet?
", "votes": null, "replies": [] }, { "id": 869965, "author_name": "gyanaluckydas", "author_url": "", "post_date": "06/01/2020 12:05:31", "content": "Yes..effdet with pseudo label
", "votes": null, "replies": [] }, { "id": 869974, "author_name": "shonenkov", "author_url": "", "post_date": "06/01/2020 12:13:50", "content": "@gyanaluckydas pseudo labeling approach can shift confidences of models, for this case should run optimizing again
", "votes": null, "replies": [] }, { "id": 869998, "author_name": "gyanaluckydas", "author_url": "", "post_date": "06/01/2020 12:22:28", "content": "I can't understand...what do u mean by optimizing again
", "votes": null, "replies": [] }, { "id": 870001, "author_name": "shonenkov", "author_url": "", "post_date": "06/01/2020 12:23:41", "content": "bayesian-optimization-wbf-efficientdet with using your pseudo labeling. you should use val fold for pseudo-labeling and holdout fold for bayesian.
\n\nWelcome!
", "votes": null, "replies": [] }, { "id": 870024, "author_name": "gyanaluckydas", "author_url": "", "post_date": "06/01/2020 12:47:08", "content": "Thanks.. I will try it
", "votes": null, "replies": [] } ] }, { "id": 875848, "author_name": "jackchungchiehyu", "author_url": "", "post_date": "06/06/2020 07:52:18", "content": "Thanks, your kernels are very instructive!
", "votes": null, "replies": [] }, { "id": 876211, "author_name": "matthewmasters", "author_url": "", "post_date": "06/06/2020 14:31:59", "content": "I've noticed that changing the iou_thr for soft_nms does not affect my accuracy at all, any idea why that could be?
Edit: Nevermind, after looking into the source code I understand now. You could remove that parameter from your soft_nms search
", "votes": null, "replies": [] }, { "id": 885459, "author_name": "cswwp347724", "author_url": "", "post_date": "06/14/2020 08:25:16", "content": "Hi @shonenkov really thank you for your kindly sharing. one question, all your mentioned scores are get with single model? such as 0.7294 --> 0.7330 cutmix
", "votes": null, "replies": [ { "id": 885470, "author_name": "shonenkov", "author_url": "", "post_date": "06/14/2020 08:43:24", "content": "@cswwp347724 thank you! this improvement for kfold
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "864688": "Hi everyone!\n\nI would like to share with you [new kernel about optimizing WBF params for EfficientDet](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet)! \n\nWhy it may be useful for you? It gives really good boost without additional training.\n\nI have got new scores:\n- 0.7176 --> 0.7193 [single](https://www.kaggle.com/shonenkov/inference-efficientdet)\n- 0.7294 --> 0.7330 [cutmix](https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet?scriptVersionId=34505469) --> [new](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet?scriptVersionId=34928761)\n- 0.7318 --> 0.7338 [cutmix + mixup](https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet?scriptVersionId=34656263)\n- 0.7375 (ensemble of two above) \n\nYou should change values of this params in your inference:\n\n```\n------------------------------\n[Best Iou Thr]: 0.44\n[Best Skip Box Thr]: 0.43\n[Best Score]: 0.6953\n------------------------------\n```\n\nWrite in comment have got boost or not ;)\n\nP.S. Also I would like to ask you, does it help anyone? Or everyone knows and uses any other values of this parameters? Do you want to know about improvements from me or I need stop with publishing? Thank you for answer in advance! \n\n[EDITED]:\n\nI have updated new version of kernel with optimization parameters for 4 methods: nms, soft_nms, nmw, wbf.", "864845": "I tried your train code, cant get that score...\nthanks for sharing, you are really imporving this compete.", "867435": "Great Kernel!\nLearning a lot from this competition. \nThanks:)", "867444": "welcome! 👍", "869844": "Very helpful kernel.\n\nI wonder whether you train from scratch or fine-tune on each fold. I achieved a small boost but maybe i am missing the actual thing.", "869869": "I don't know why this doesn't work for me\nBefore Bayesian Optimization : 0.7355\nAfter Bayesian Optimization : 0.7346", "869933": "gyanaluckydas what models did you use? effdet?", "869965": "Yes..effdet with pseudo label", "869974": "gyanaluckydas pseudo labeling approach can shift confidences of models, for this case should run optimizing again", "869998": "I can't understand...what do u mean by optimizing again", "870001": "[bayesian-optimization-wbf-efficientdet](https://www.kaggle.com/shonenkov/bayesian-optimization-wbf-efficientdet) with using your pseudo labeling. you should use val fold for pseudo-labeling and holdout fold for bayesian.\n\nWelcome!", "870024": "Thanks.. I will try it", "875848": "Thanks, your kernels are very instructive!", "876211": "I've noticed that changing the `iou_thr` for `soft_nms` does not affect my accuracy at all, any idea why that could be?\n\nEdit: Nevermind, after looking into the source code I understand now. You could remove that parameter from your soft_nms search", "885459": "Hi @shonenkov really thank you for your kindly sharing. one question, all your mentioned scores are get with single model? such as 0.7294 --> 0.7330 cutmix", "885470": "cswwp347724 thank you! this improvement for kfold" }, "source": "meta" }