I have used Keras trained model (Xception, Densenet, resnet50) and the latest EfficientNet, and would like to share the LB score of them:
\n\nXception: 0.677 (rmsprop optimiser by default )\nResnet 50: 0.708 (adam)\nDensenet: 0.729 (adam)\nEfficientNet: 0.726 (adam)
\n\nI have made the EfficientNet kernel available as well if you are interested, please find it here\nEfficientNet.
\n\nWould you like to share your performance of different models?
", "messages": [ { "id": "578315", "postDate": "07/17/2019 15:44:36", "content": "I have used Keras trained model (Xception, Densenet, resnet50) and the latest EfficientNet, and would like to share the LB score of them:
\n\nXception: 0.677 (rmsprop optimiser by default )\nResnet 50: 0.708 (adam)\nDensenet: 0.729 (adam)\nEfficientNet: 0.726 (adam)
\n\nI have made the EfficientNet kernel available as well if you are interested, please find it here\nEfficientNet.
\n\nWould you like to share your performance of different models?
", "rawMarkdown": "I have used Keras trained model (Xception, Densenet, resnet50) and the latest [EfficientNet](https://www.kaggle.com/runninglion/aptos-2019-efficientnet-keras-model-weights), and would like to share the LB score of them:\n\nXception: 0.677 (rmsprop optimiser by default )\nResnet 50: 0.708 (adam)\nDensenet: 0.729 (adam)\nEfficientNet: 0.726 (adam)\n\nI have made the EfficientNet kernel available as well if you are interested, please find it here\n[EfficientNet](https://www.kaggle.com/runninglion/aptos-2019-efficientnet-keras-model-weights).\n\nWould you like to share your performance of different models?", "votes": null }, { "id": "578344", "postDate": "07/17/2019 16:20:47", "content": "Resnext101_32x4d: 0.767\nModel performance depends on a lot of other factors too.
", "rawMarkdown": "Resnext101_32x4d: 0.767\nModel performance depends on a lot of other factors too.", "votes": null }, { "id": "578549", "postDate": "07/17/2019 20:57:50", "content": "Yes, definitely and thank you for sharing your result. I posted those performances without using data augmentation, data preprocessing and others. I am keen to see someone can share their performances considering other factors as well.
", "rawMarkdown": "Yes, definitely and thank you for sharing your result. I posted those performances without using data augmentation, data preprocessing and others. I am keen to see someone can share their performances considering other factors as well.", "votes": null }, { "id": "578601", "postDate": "07/17/2019 22:52:25", "content": "My Xception model gives 0.699 Lb
", "rawMarkdown": "My Xception model gives 0.699 Lb", "votes": null }, { "id": "578658", "postDate": "07/18/2019 02:06:27", "content": "EfficientNetB0 with LB 0.77
", "rawMarkdown": "EfficientNetB0 with LB 0.77", "votes": null }, { "id": "579500", "postDate": "07/18/2019 22:30:43", "content": "thank you for sharing your result, do you use classification or regression? if you have tried both, which one is better?
", "rawMarkdown": "thank you for sharing your result, do you use classification or regression? if you have tried both, which one is better?", "votes": null }, { "id": "579501", "postDate": "07/18/2019 22:31:24", "content": "thank you for sharing your result, have you tried other models?
", "rawMarkdown": "thank you for sharing your result, have you tried other models?", "votes": null }, { "id": "580086", "postDate": "07/19/2019 16:46:19", "content": "DenseNet: 0.752\nResNet: 0.721\nEfficientNet: 0.726
\n\nThe thing is that, IMHO, this comparison is not truly insightful. I'm more and more concerned that we might rely totally on CV scores rather than Public LB.
\n\nMy best validation scores were around 0.88 with EfficientNet. Unfortunately when I submit that I got around 0.7 LB.\nTo conclude, my predictions have a predominance of 2's and 3's while a higher score has more 2's and less 3's.
", "rawMarkdown": "DenseNet: 0.752\nResNet: 0.721\nEfficientNet: 0.726\n\nThe thing is that, IMHO, this comparison is not truly insightful. I'm more and more concerned that we might rely totally on CV scores rather than Public LB.\n\nMy best validation scores were around 0.88 with EfficientNet. Unfortunately when I submit that I got around 0.7 LB.\nTo conclude, my predictions have a predominance of 2's and 3's while a higher score has more 2's and less 3's.", "votes": null }, { "id": "580294", "postDate": "07/20/2019 01:23:38", "content": "Because of the imbalanced data, I think it must be regression.
", "rawMarkdown": "Because of the imbalanced data, I think it must be regression.", "votes": null }, { "id": "580297", "postDate": "07/20/2019 01:29:01", "content": "I don't know what happens to you , but EfficientNet should worth more LB score
", "rawMarkdown": "I don't know what happens to you , but EfficientNet should worth more LB score", "votes": null }, { "id": "616232", "postDate": "09/02/2019 21:52:57", "content": "I used densenet with Adam and got LB 0.765. I am planning to use 2015 competition images as training images and see if that improves the score.
", "rawMarkdown": "I used densenet with Adam and got LB 0.765. I am planning to use 2015 competition images as training images and see if that improves the score.", "votes": null }, { "id": "616235", "postDate": "09/02/2019 22:02:43", "content": "@xujingzhao Actually, right now the best model is EffNet indeed. Fine-tuning on new competition data helped me to pass from around 0.75 to 0.78.
", "rawMarkdown": "xujingzhao Actually, right now the best model is EffNet indeed. Fine-tuning on new competition data helped me to pass from around 0.75 to 0.78.", "votes": null }, { "id": "616240", "postDate": "09/02/2019 22:13:08", "content": "@raimonds1993 : That's a good improvement! Did you use EfficientNetB5?
", "rawMarkdown": "raimonds1993 : That's a good improvement! Did you use EfficientNetB5?", "votes": null }, { "id": "618151", "postDate": "09/04/2019 23:37:32", "content": "Resnet 152 0.737
", "rawMarkdown": "Resnet 152 0.737", "votes": null } ], "comments": [ { "id": 578344, "author_name": "rishabhiitbhu", "author_url": "", "post_date": "07/17/2019 16:20:47", "content": "Resnext101_32x4d: 0.767\nModel performance depends on a lot of other factors too.
", "votes": null, "replies": [ { "id": 578549, "author_name": "runninglion", "author_url": "", "post_date": "07/17/2019 20:57:50", "content": "Yes, definitely and thank you for sharing your result. I posted those performances without using data augmentation, data preprocessing and others. I am keen to see someone can share their performances considering other factors as well.
", "votes": null, "replies": [] } ] }, { "id": 578601, "author_name": "jmourad100", "author_url": "", "post_date": "07/17/2019 22:52:25", "content": "My Xception model gives 0.699 Lb
", "votes": null, "replies": [ { "id": 579501, "author_name": "runninglion", "author_url": "", "post_date": "07/18/2019 22:31:24", "content": "thank you for sharing your result, have you tried other models?
", "votes": null, "replies": [] } ] }, { "id": 578658, "author_name": "xujingzhao", "author_url": "", "post_date": "07/18/2019 02:06:27", "content": "EfficientNetB0 with LB 0.77
", "votes": null, "replies": [ { "id": 579500, "author_name": "runninglion", "author_url": "", "post_date": "07/18/2019 22:30:43", "content": "thank you for sharing your result, do you use classification or regression? if you have tried both, which one is better?
", "votes": null, "replies": [] }, { "id": 580294, "author_name": "xujingzhao", "author_url": "", "post_date": "07/20/2019 01:23:38", "content": "Because of the imbalanced data, I think it must be regression.
", "votes": null, "replies": [] } ] }, { "id": 580086, "author_name": "raimonds1993", "author_url": "", "post_date": "07/19/2019 16:46:19", "content": "DenseNet: 0.752\nResNet: 0.721\nEfficientNet: 0.726
\n\nThe thing is that, IMHO, this comparison is not truly insightful. I'm more and more concerned that we might rely totally on CV scores rather than Public LB.
\n\nMy best validation scores were around 0.88 with EfficientNet. Unfortunately when I submit that I got around 0.7 LB.\nTo conclude, my predictions have a predominance of 2's and 3's while a higher score has more 2's and less 3's.
", "votes": null, "replies": [ { "id": 580297, "author_name": "xujingzhao", "author_url": "", "post_date": "07/20/2019 01:29:01", "content": "I don't know what happens to you , but EfficientNet should worth more LB score
", "votes": null, "replies": [] }, { "id": 616235, "author_name": "raimonds1993", "author_url": "", "post_date": "09/02/2019 22:02:43", "content": "@xujingzhao Actually, right now the best model is EffNet indeed. Fine-tuning on new competition data helped me to pass from around 0.75 to 0.78.
", "votes": null, "replies": [] }, { "id": 616240, "author_name": "rmuthiah", "author_url": "", "post_date": "09/02/2019 22:13:08", "content": "@raimonds1993 : That's a good improvement! Did you use EfficientNetB5?
", "votes": null, "replies": [] } ] }, { "id": 616232, "author_name": "rmuthiah", "author_url": "", "post_date": "09/02/2019 21:52:57", "content": "I used densenet with Adam and got LB 0.765. I am planning to use 2015 competition images as training images and see if that improves the score.
", "votes": null, "replies": [] }, { "id": 618151, "author_name": "chrisfs", "author_url": "", "post_date": "09/04/2019 23:37:32", "content": "Resnet 152 0.737
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "578315": "I have used Keras trained model (Xception, Densenet, resnet50) and the latest [EfficientNet](https://www.kaggle.com/runninglion/aptos-2019-efficientnet-keras-model-weights), and would like to share the LB score of them:\n\nXception: 0.677 (rmsprop optimiser by default )\nResnet 50: 0.708 (adam)\nDensenet: 0.729 (adam)\nEfficientNet: 0.726 (adam)\n\nI have made the EfficientNet kernel available as well if you are interested, please find it here\n[EfficientNet](https://www.kaggle.com/runninglion/aptos-2019-efficientnet-keras-model-weights).\n\nWould you like to share your performance of different models?", "578344": "Resnext101_32x4d: 0.767\nModel performance depends on a lot of other factors too.", "578549": "Yes, definitely and thank you for sharing your result. I posted those performances without using data augmentation, data preprocessing and others. I am keen to see someone can share their performances considering other factors as well.", "578601": "My Xception model gives 0.699 Lb", "578658": "EfficientNetB0 with LB 0.77", "579500": "thank you for sharing your result, do you use classification or regression? if you have tried both, which one is better?", "579501": "thank you for sharing your result, have you tried other models?", "580086": "DenseNet: 0.752\nResNet: 0.721\nEfficientNet: 0.726\n\nThe thing is that, IMHO, this comparison is not truly insightful. I'm more and more concerned that we might rely totally on CV scores rather than Public LB.\n\nMy best validation scores were around 0.88 with EfficientNet. Unfortunately when I submit that I got around 0.7 LB.\nTo conclude, my predictions have a predominance of 2's and 3's while a higher score has more 2's and less 3's.", "580294": "Because of the imbalanced data, I think it must be regression.", "580297": "I don't know what happens to you , but EfficientNet should worth more LB score", "616232": "I used densenet with Adam and got LB 0.765. I am planning to use 2015 competition images as training images and see if that improves the score.", "616235": "xujingzhao Actually, right now the best model is EffNet indeed. Fine-tuning on new competition data helped me to pass from around 0.75 to 0.78.", "616240": "raimonds1993 : That's a good improvement! Did you use EfficientNetB5?", "618151": "Resnet 152 0.737" }, "source": "meta" }