i am currently using efficientdet-5\nlb = 0.7445 (no pseudo labeling)\nwith pseudo labeling lb become 0.7323\nis anyone facing this problem or have solved this problem\nplease give some tips
", "messages": [ { "id": "930253", "postDate": "07/15/2020 09:56:40", "content": "i am currently using efficientdet-5\nlb = 0.7445 (no pseudo labeling)\nwith pseudo labeling lb become 0.7323\nis anyone facing this problem or have solved this problem\nplease give some tips
", "rawMarkdown": "i am currently using efficientdet-5\nlb = 0.7445 (no pseudo labeling)\nwith pseudo labeling lb become 0.7323\nis anyone facing this problem or have solved this problem\nplease give some tips", "votes": null }, { "id": "930259", "postDate": "07/15/2020 10:01:05", "content": "Does efficientdet comply with competition rules? Isnt it Apache licensed?
", "rawMarkdown": "Does efficientdet comply with competition rules? Isnt it Apache licensed?", "votes": null }, { "id": "930304", "postDate": "07/15/2020 10:50:18", "content": "@abhishek I came across this https://github.com/rwightman/efficientdet-pytorch/issues/51 .. also here https://www.kaggle.com/c/global-wheat-detection/discussion/164845#923524
", "rawMarkdown": "abhishek I came across this https://github.com/rwightman/efficientdet-pytorch/issues/51 .. also here https://www.kaggle.com/c/global-wheat-detection/discussion/164845#923524", "votes": null }, { "id": "930565", "postDate": "07/15/2020 14:51:18", "content": "I don't use efficientDet, but I also faced with the problem of decreasing LB scores after using pseudo labeling.
", "rawMarkdown": "I don't use efficientDet, but I also faced with the problem of decreasing LB scores after using pseudo labeling.", "votes": null }, { "id": "930844", "postDate": "07/15/2020 18:41:23", "content": "Maybe you are overfitting? Have you selected only the most confident bounding boxes?
", "rawMarkdown": "Maybe you are overfitting? Have you selected only the most confident bounding boxes?", "votes": null }, { "id": "930860", "postDate": "07/15/2020 18:53:14", "content": "yeah with some tta and wbf
", "rawMarkdown": "yeah with some tta and wbf", "votes": null }, { "id": "930865", "postDate": "07/15/2020 18:56:51", "content": "Are you using the TTA for both the pseudolabels and final prediction? If you left it off of one, maybe that's the reason for the lower score(?)
", "rawMarkdown": "Are you using the TTA for both the pseudolabels and final prediction? If you left it off of one, maybe that's the reason for the lower score(?)", "votes": null }, { "id": "930877", "postDate": "07/15/2020 19:14:22", "content": "tta and wbf for both the pseudolabels and final prediction
", "rawMarkdown": "tta and wbf for both the pseudolabels and final prediction", "votes": null }, { "id": "931013", "postDate": "07/15/2020 22:36:05", "content": "I have the same kind of problem. Did you figure out why? I guess it's just overfitting?
", "rawMarkdown": "I have the same kind of problem. Did you figure out why? I guess it's just overfitting?", "votes": null }, { "id": "935801", "postDate": "07/19/2020 16:34:15", "content": "https://www.kaggle.com/c/global-wheat-detection/discussion/167298
\nThere are many commonly used permissive open source licenses (e.g., Apache 2.0, BSD 3.0, others) that are compatible with the MIT License and therefore are compliant with the Competition Rules.
", "rawMarkdown": "https://www.kaggle.com/c/global-wheat-detection/discussion/167298\n\n\nThere are many commonly used permissive open source licenses (e.g., Apache 2.0, BSD 3.0, others) that are compatible with the MIT License and therefore are compliant with the Competition Rules.", "votes": null }, { "id": "936779", "postDate": "07/20/2020 14:19:58", "content": "Have you solved this? I have the same problem.
", "rawMarkdown": "Have you solved this? I have the same problem.", "votes": null }, { "id": "936797", "postDate": "07/20/2020 14:31:24", "content": "i have used very strong augmentation
\ncutmix and affine in pseudo labeling
\nit increases lb score
Cool ! Thanks
", "rawMarkdown": "Cool ! Thanks", "votes": null }, { "id": "936868", "postDate": "07/20/2020 15:04:52", "content": "That's great! Have you trained using this notebook: https://www.kaggle.com/shonenkov/training-efficientdet? If so, how have you made it work? It seems the input of the net = get_net() model has changed a bit... \nThanks :)
", "rawMarkdown": "That's great! Have you trained using this notebook: https://www.kaggle.com/shonenkov/training-efficientdet? If so, how have you made it work? It seems the input of the net = get_net() model has changed a bit... \nThanks :)", "votes": null }, { "id": "936983", "postDate": "07/20/2020 16:28:48", "content": "https://www.kaggle.com/sramirez/training-efficientdet-gradient-accumulation-f1
\nhttps://www.kaggle.com/shonenkov/training-efficientdet
", "rawMarkdown": "https://www.kaggle.com/sramirez/training-efficientdet-gradient-accumulation-f1\n\nhttps://www.kaggle.com/shonenkov/training-efficientdet", "votes": null }, { "id": "937058", "postDate": "07/20/2020 17:40:23", "content": "Thanks for the first notebook. It seems the second input changed in the following fashion:
\n\n```\n targets2 = {}\n targets2['bbox'] = [target['boxes'].to(self.device).float() for target in targets] # variable number of instances, so the entire structure can be forced to tensor\n targets2['cls'] = [target['labels'].to(self.device).float() for target in targets]\n targets2['image_id'] = torch.tensor([target['image_id'] for target in targets]).to(self.device).float()\n targets2['img_scale'] = torch.tensor([target['img_scale'] for target in targets]).to(self.device).float()\n targets2['img_size'] = torch.tensor([(IMG_SIZE, IMG_SIZE) for target in targets]).to(self.device).float()
\n\n output = self.model(images, targets2)\n```
", "rawMarkdown": "Thanks for the first notebook. It seems the second input changed in the following fashion: \n\n```\n targets2 = {}\n targets2['bbox'] = [target['boxes'].to(self.device).float() for target in targets] # variable number of instances, so the entire structure can be forced to tensor\n targets2['cls'] = [target['labels'].to(self.device).float() for target in targets]\n targets2['image_id'] = torch.tensor([target['image_id'] for target in targets]).to(self.device).float()\n targets2['img_scale'] = torch.tensor([target['img_scale'] for target in targets]).to(self.device).float()\n targets2['img_size'] = torch.tensor([(IMG_SIZE, IMG_SIZE) for target in targets]).to(self.device).float()\n \n output = self.model(images, targets2)\n```", "votes": null } ], "comments": [ { "id": 935801, "author_name": "ajax0564", "author_url": "", "post_date": "07/19/2020 16:34:15", "content": "https://www.kaggle.com/c/global-wheat-detection/discussion/167298
\nThere are many commonly used permissive open source licenses (e.g., Apache 2.0, BSD 3.0, others) that are compatible with the MIT License and therefore are compliant with the Competition Rules.
", "votes": null, "replies": [] }, { "id": 930259, "author_name": "abhishek", "author_url": "", "post_date": "07/15/2020 10:01:05", "content": "Does efficientdet comply with competition rules? Isnt it Apache licensed?
", "votes": null, "replies": [ { "id": 930304, "author_name": "kalyanpichuka", "author_url": "", "post_date": "07/15/2020 10:50:18", "content": "@abhishek I came across this https://github.com/rwightman/efficientdet-pytorch/issues/51 .. also here https://www.kaggle.com/c/global-wheat-detection/discussion/164845#923524
", "votes": null, "replies": [] } ] }, { "id": 930565, "author_name": "solosquad1999", "author_url": "", "post_date": "07/15/2020 14:51:18", "content": "I don't use efficientDet, but I also faced with the problem of decreasing LB scores after using pseudo labeling.
", "votes": null, "replies": [] }, { "id": 930844, "author_name": "yassinealouini", "author_url": "", "post_date": "07/15/2020 18:41:23", "content": "Maybe you are overfitting? Have you selected only the most confident bounding boxes?
", "votes": null, "replies": [ { "id": 930860, "author_name": "ajax0564", "author_url": "", "post_date": "07/15/2020 18:53:14", "content": "yeah with some tta and wbf
", "votes": null, "replies": [] }, { "id": 930865, "author_name": "matthewmasters", "author_url": "", "post_date": "07/15/2020 18:56:51", "content": "Are you using the TTA for both the pseudolabels and final prediction? If you left it off of one, maybe that's the reason for the lower score(?)
", "votes": null, "replies": [] }, { "id": 930877, "author_name": "ajax0564", "author_url": "", "post_date": "07/15/2020 19:14:22", "content": "tta and wbf for both the pseudolabels and final prediction
", "votes": null, "replies": [] } ] }, { "id": 931013, "author_name": "jonykarki", "author_url": "", "post_date": "07/15/2020 22:36:05", "content": "I have the same kind of problem. Did you figure out why? I guess it's just overfitting?
", "votes": null, "replies": [] }, { "id": 936779, "author_name": "sqqqqy", "author_url": "", "post_date": "07/20/2020 14:19:58", "content": "Have you solved this? I have the same problem.
", "votes": null, "replies": [ { "id": 936797, "author_name": "ajax0564", "author_url": "", "post_date": "07/20/2020 14:31:24", "content": "i have used very strong augmentation
\ncutmix and affine in pseudo labeling
\nit increases lb score
Cool ! Thanks
", "votes": null, "replies": [] }, { "id": 936868, "author_name": "yassinealouini", "author_url": "", "post_date": "07/20/2020 15:04:52", "content": "That's great! Have you trained using this notebook: https://www.kaggle.com/shonenkov/training-efficientdet? If so, how have you made it work? It seems the input of the net = get_net() model has changed a bit... \nThanks :)
", "votes": null, "replies": [] }, { "id": 936983, "author_name": "ajax0564", "author_url": "", "post_date": "07/20/2020 16:28:48", "content": "https://www.kaggle.com/sramirez/training-efficientdet-gradient-accumulation-f1
\nhttps://www.kaggle.com/shonenkov/training-efficientdet
", "votes": null, "replies": [] }, { "id": 937058, "author_name": "yassinealouini", "author_url": "", "post_date": "07/20/2020 17:40:23", "content": "Thanks for the first notebook. It seems the second input changed in the following fashion:
\n\n```\n targets2 = {}\n targets2['bbox'] = [target['boxes'].to(self.device).float() for target in targets] # variable number of instances, so the entire structure can be forced to tensor\n targets2['cls'] = [target['labels'].to(self.device).float() for target in targets]\n targets2['image_id'] = torch.tensor([target['image_id'] for target in targets]).to(self.device).float()\n targets2['img_scale'] = torch.tensor([target['img_scale'] for target in targets]).to(self.device).float()\n targets2['img_size'] = torch.tensor([(IMG_SIZE, IMG_SIZE) for target in targets]).to(self.device).float()
\n\n output = self.model(images, targets2)\n```
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "930253": "i am currently using efficientdet-5\nlb = 0.7445 (no pseudo labeling)\nwith pseudo labeling lb become 0.7323\nis anyone facing this problem or have solved this problem\nplease give some tips", "930259": "Does efficientdet comply with competition rules? Isnt it Apache licensed?", "930304": "abhishek I came across this https://github.com/rwightman/efficientdet-pytorch/issues/51 .. also here https://www.kaggle.com/c/global-wheat-detection/discussion/164845#923524", "930565": "I don't use efficientDet, but I also faced with the problem of decreasing LB scores after using pseudo labeling.", "930844": "Maybe you are overfitting? Have you selected only the most confident bounding boxes?", "930860": "yeah with some tta and wbf", "930865": "Are you using the TTA for both the pseudolabels and final prediction? If you left it off of one, maybe that's the reason for the lower score(?)", "930877": "tta and wbf for both the pseudolabels and final prediction", "931013": "I have the same kind of problem. Did you figure out why? I guess it's just overfitting?", "935801": "https://www.kaggle.com/c/global-wheat-detection/discussion/167298\n\n\nThere are many commonly used permissive open source licenses (e.g., Apache 2.0, BSD 3.0, others) that are compatible with the MIT License and therefore are compliant with the Competition Rules.", "936779": "Have you solved this? I have the same problem.", "936797": "i have used very strong augmentation\ncutmix and affine in pseudo labeling\nit increases lb score", "936802": "Cool ! Thanks", "936868": "That's great! Have you trained using this notebook: https://www.kaggle.com/shonenkov/training-efficientdet? If so, how have you made it work? It seems the input of the net = get_net() model has changed a bit... \nThanks :)", "936983": "https://www.kaggle.com/sramirez/training-efficientdet-gradient-accumulation-f1\n\nhttps://www.kaggle.com/shonenkov/training-efficientdet", "937058": "Thanks for the first notebook. It seems the second input changed in the following fashion: \n\n```\n targets2 = {}\n targets2['bbox'] = [target['boxes'].to(self.device).float() for target in targets] # variable number of instances, so the entire structure can be forced to tensor\n targets2['cls'] = [target['labels'].to(self.device).float() for target in targets]\n targets2['image_id'] = torch.tensor([target['image_id'] for target in targets]).to(self.device).float()\n targets2['img_scale'] = torch.tensor([target['img_scale'] for target in targets]).to(self.device).float()\n targets2['img_size'] = torch.tensor([(IMG_SIZE, IMG_SIZE) for target in targets]).to(self.device).float()\n \n output = self.model(images, targets2)\n```" }, "source": "meta" }