I am new enterant into this competition.\nCan any one please point me to formidable train pipeline compliant with competition rules.
", "messages": [ { "id": "931324", "postDate": "07/16/2020 05:54:36", "content": "I am new enterant into this competition.\nCan any one please point me to formidable train pipeline compliant with competition rules.
", "rawMarkdown": "I am new enterant into this competition.\nCan any one please point me to formidable train pipeline compliant with competition rules.", "votes": null }, { "id": "931332", "postDate": "07/16/2020 06:02:14", "content": "The big issue was with YoloV5 since it's licenced under GPLv3. You can take a look here for other frameworks and their associated licence, EffDet and YoloV3 seem to be popular choices. Most people are using WBF for ensembling which is now under MIT licence. I'd recommend you look at some of the existing public notebooks to get an idea on how your pipeline should work.
", "rawMarkdown": "The big issue was with YoloV5 since it's licenced under GPLv3. You can take a look [here](https://www.kaggle.com/c/global-wheat-detection/discussion/165735#924609) for other frameworks and their associated licence, EffDet and YoloV3 seem to be popular choices. Most people are using [WBF](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) for ensembling which is now under MIT licence. I'd recommend you look at some of the existing public notebooks to get an idea on how your pipeline should work.", "votes": null }, { "id": "931334", "postDate": "07/16/2020 06:04:38", "content": "thanks ditn any one tried retina net . owing to different size of bbox we have got for the labels ?
", "rawMarkdown": "thanks ditn any one tried retina net . owing to different size of bbox we have got for the labels ?", "votes": null }, { "id": "931346", "postDate": "07/16/2020 06:16:23", "content": "Yeah, someone tried it here, though it only got 0.6 on the LB. I also thought it might be a good approach but never got around to trying it out.
", "rawMarkdown": "Yeah, someone tried it [here](https://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission), though it only got 0.6 on the LB. I also thought it might be a good approach but never got around to trying it out.", "votes": null }, { "id": "931350", "postDate": "07/16/2020 06:25:40", "content": "ok OOF is same as WBF ?\n1) what it stands for ?\n2) From where i can get i can get validation metrics ,is it IOU or mAP
", "rawMarkdown": "ok OOF is same as WBF ?\n1) what it stands for ?\n2) From where i can get i can get validation metrics ,is it IOU or mAP", "votes": null }, { "id": "931362", "postDate": "07/16/2020 06:37:53", "content": "Yeah, you can use WBF to ensemble k-fold models and make your predictions OOF.\n- WBF stands for weighted box fusion. There are other options such as NMS, Soft-NMS, etc. but WBF seems to work well for this challenge. It was invented by a fellow Kaggler @zfturbo and you can find the paper describing it here.\n- The validation metric is mean average precision calculated from 0.5 IOU to 0.75 in 0.05 increments. @pestipeti wrote a great implementation of this metric which can be found here.
\n\nP.S. the search tool is your friend 😉
", "rawMarkdown": "Yeah, you can use WBF to ensemble k-fold models and make your predictions OOF.\n- WBF stands for weighted box fusion. There are other options such as NMS, Soft-NMS, etc. but WBF seems to work well for this challenge. It was invented by a fellow Kaggler @zfturbo and you can find the paper describing it [here](https://arxiv.org/pdf/1910.13302.pdf).\n- The validation metric is mean average precision calculated from 0.5 IOU to 0.75 in 0.05 increments. @pestipeti wrote a great implementation of this metric which can be found [here](https://www.kaggle.com/pestipeti/competition-metric-details-script).\n\nP.S. the search tool is your friend 😉", "votes": null }, { "id": "931387", "postDate": "07/16/2020 07:02:01", "content": "try this: https://www.kaggle.com/shonenkov/training-efficientdet
", "rawMarkdown": "try this: https://www.kaggle.com/shonenkov/training-efficientdet", "votes": null } ], "comments": [ { "id": 931332, "author_name": "matthewmasters", "author_url": "", "post_date": "07/16/2020 06:02:14", "content": "The big issue was with YoloV5 since it's licenced under GPLv3. You can take a look here for other frameworks and their associated licence, EffDet and YoloV3 seem to be popular choices. Most people are using WBF for ensembling which is now under MIT licence. I'd recommend you look at some of the existing public notebooks to get an idea on how your pipeline should work.
", "votes": null, "replies": [ { "id": 931334, "author_name": "jaideepvalani", "author_url": "", "post_date": "07/16/2020 06:04:38", "content": "thanks ditn any one tried retina net . owing to different size of bbox we have got for the labels ?
", "votes": null, "replies": [] }, { "id": 931346, "author_name": "matthewmasters", "author_url": "", "post_date": "07/16/2020 06:16:23", "content": "Yeah, someone tried it here, though it only got 0.6 on the LB. I also thought it might be a good approach but never got around to trying it out.
", "votes": null, "replies": [] }, { "id": 931350, "author_name": "jaideepvalani", "author_url": "", "post_date": "07/16/2020 06:25:40", "content": "ok OOF is same as WBF ?\n1) what it stands for ?\n2) From where i can get i can get validation metrics ,is it IOU or mAP
", "votes": null, "replies": [] }, { "id": 931362, "author_name": "matthewmasters", "author_url": "", "post_date": "07/16/2020 06:37:53", "content": "Yeah, you can use WBF to ensemble k-fold models and make your predictions OOF.\n- WBF stands for weighted box fusion. There are other options such as NMS, Soft-NMS, etc. but WBF seems to work well for this challenge. It was invented by a fellow Kaggler @zfturbo and you can find the paper describing it here.\n- The validation metric is mean average precision calculated from 0.5 IOU to 0.75 in 0.05 increments. @pestipeti wrote a great implementation of this metric which can be found here.
\n\nP.S. the search tool is your friend 😉
", "votes": null, "replies": [] } ] }, { "id": 931387, "author_name": "tiandaye", "author_url": "", "post_date": "07/16/2020 07:02:01", "content": "try this: https://www.kaggle.com/shonenkov/training-efficientdet
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "931324": "I am new enterant into this competition.\nCan any one please point me to formidable train pipeline compliant with competition rules.", "931332": "The big issue was with YoloV5 since it's licenced under GPLv3. You can take a look [here](https://www.kaggle.com/c/global-wheat-detection/discussion/165735#924609) for other frameworks and their associated licence, EffDet and YoloV3 seem to be popular choices. Most people are using [WBF](https://github.com/ZFTurbo/Weighted-Boxes-Fusion) for ensembling which is now under MIT licence. I'd recommend you look at some of the existing public notebooks to get an idea on how your pipeline should work.", "931334": "thanks ditn any one tried retina net . owing to different size of bbox we have got for the labels ?", "931346": "Yeah, someone tried it [here](https://www.kaggle.com/labintsevai/kerasretinanet-wheat-submission), though it only got 0.6 on the LB. I also thought it might be a good approach but never got around to trying it out.", "931350": "ok OOF is same as WBF ?\n1) what it stands for ?\n2) From where i can get i can get validation metrics ,is it IOU or mAP", "931362": "Yeah, you can use WBF to ensemble k-fold models and make your predictions OOF.\n- WBF stands for weighted box fusion. There are other options such as NMS, Soft-NMS, etc. but WBF seems to work well for this challenge. It was invented by a fellow Kaggler @zfturbo and you can find the paper describing it [here](https://arxiv.org/pdf/1910.13302.pdf).\n- The validation metric is mean average precision calculated from 0.5 IOU to 0.75 in 0.05 increments. @pestipeti wrote a great implementation of this metric which can be found [here](https://www.kaggle.com/pestipeti/competition-metric-details-script).\n\nP.S. the search tool is your friend 😉", "931387": "try this: https://www.kaggle.com/shonenkov/training-efficientdet" }, "source": "meta" }