Congrats to all the winners! Actually I expected to have bigger shakeup in the private LB but it seems that trusting the public LB was a good idea. I selected my top 2 LB and it worked : \n- 0.82 LB and in the private it gave 0.92 kappa \n- 0.817 LB and 0.92 in the private too\nThese both kernels are an ensemble of 4 models : \n- EfficientNet b0 , size=416 (btw this single model gave 0.806 in LB and 0.92 in the private)\n- EfficientNet b3 , size=288 (0.805 LB vs 0.913 private)\n- EfficientNet b4 , size=288 (did not submit it)\n- EfficientNet b5 , size=256 (0.797 LB vs 0.914 private)\nThe only difference between my top 2 kernels was the TTA:\nI just used a simple 5-TTA with these augmentations : \n \nalbumentations.Compose([\n albumentations.Resize(S, S),\n albumentations.ShiftScaleRotate(shift_limit=0.08, scale_limit=0.15, rotate_limit=30, p=0.5),\n albumentations.Normalize(),\n AT.ToTensor(),\n ])\n
but gave me worse public LB. I thought TTA would help in the private LB but it seems that in all my TTA kernels scores went down even in private LB.\nAnother thing to add, all my models were trained on 80% of the data. I did not use Kfolds since Kfolds gave me bad results and I think I have chosen the golden fold that gave me more 0.02 score than the other folds. That's why ensembling did'nt gave me a huge boost since I am only validating my models on ~700 images.\nThank you for everyone who shared a kernel or an idea or made a comment or a post in the discussion section. \nThank you Kaggle and APTOS for this amazing competition.
\n\n[Update : I trained every model on the old data of the previous competition (train+test) using the current data as validation. All my models converge to 0.88 kappa with pretrained weights]
", "messages": [ { "id": "620768", "postDate": "09/08/2019 00:23:07", "content": "Congrats to all the winners! Actually I expected to have bigger shakeup in the private LB but it seems that trusting the public LB was a good idea. I selected my top 2 LB and it worked : \n- 0.82 LB and in the private it gave 0.92 kappa \n- 0.817 LB and 0.92 in the private too\nThese both kernels are an ensemble of 4 models : \n- EfficientNet b0 , size=416 (btw this single model gave 0.806 in LB and 0.92 in the private)\n- EfficientNet b3 , size=288 (0.805 LB vs 0.913 private)\n- EfficientNet b4 , size=288 (did not submit it)\n- EfficientNet b5 , size=256 (0.797 LB vs 0.914 private)\nThe only difference between my top 2 kernels was the TTA:\nI just used a simple 5-TTA with these augmentations : \n \nalbumentations.Compose([\n albumentations.Resize(S, S),\n albumentations.ShiftScaleRotate(shift_limit=0.08, scale_limit=0.15, rotate_limit=30, p=0.5),\n albumentations.Normalize(),\n AT.ToTensor(),\n ])\n
but gave me worse public LB. I thought TTA would help in the private LB but it seems that in all my TTA kernels scores went down even in private LB.\nAnother thing to add, all my models were trained on 80% of the data. I did not use Kfolds since Kfolds gave me bad results and I think I have chosen the golden fold that gave me more 0.02 score than the other folds. That's why ensembling did'nt gave me a huge boost since I am only validating my models on ~700 images.\nThank you for everyone who shared a kernel or an idea or made a comment or a post in the discussion section. \nThank you Kaggle and APTOS for this amazing competition.
\n\n[Update : I trained every model on the old data of the previous competition (train+test) using the current data as validation. All my models converge to 0.88 kappa with pretrained weights]
", "rawMarkdown": "Congrats to all the winners! Actually I expected to have bigger shakeup in the private LB but it seems that trusting the public LB was a good idea. I selected my top 2 LB and it worked : \n- 0.82 LB and in the private it gave 0.92 kappa \n- 0.817 LB and 0.92 in the private too\nThese both kernels are an ensemble of 4 models : \n- EfficientNet b0 , size=416 (btw this single model gave 0.806 in LB and 0.92 in the private)\n- EfficientNet b3 , size=288 (0.805 LB vs 0.913 private)\n- EfficientNet b4 , size=288 (did not submit it)\n- EfficientNet b5 , size=256 (0.797 LB vs 0.914 private)\nThe only difference between my top 2 kernels was the TTA:\nI just used a simple 5-TTA with these augmentations : \n ```\nalbumentations.Compose([\n albumentations.Resize(S, S),\n albumentations.ShiftScaleRotate(shift_limit=0.08, scale_limit=0.15, rotate_limit=30, p=0.5),\n albumentations.Normalize(),\n AT.ToTensor(),\n ])\n```\n\nbut gave me worse public LB. I thought TTA would help in the private LB but it seems that in all my TTA kernels scores went down even in private LB.\nAnother thing to add, all my models were trained on 80% of the data. I did not use Kfolds since Kfolds gave me bad results and I think I have chosen the golden fold that gave me more 0.02 score than the other folds. That's why ensembling did'nt gave me a huge boost since I am only validating my models on ~700 images.\nThank you for everyone who shared a kernel or an idea or made a comment or a post in the discussion section. \nThank you Kaggle and APTOS for this amazing competition.\n\n[Update : I trained every model on the old data of the previous competition (train+test) using the current data as validation. All my models converge to 0.88 kappa with pretrained weights]", "votes": null }, { "id": "620910", "postDate": "09/08/2019 04:06:22", "content": "Congratulations\nGreat work\nAnd Thanks for sharing your approach and insights.!! @rinnqd
", "rawMarkdown": "Congratulations\nGreat work\nAnd Thanks for sharing your approach and insights.!! @rinnqd", "votes": null } ], "comments": [ { "id": 620910, "author_name": "veeralakrishna", "author_url": "", "post_date": "09/08/2019 04:06:22", "content": "Congratulations\nGreat work\nAnd Thanks for sharing your approach and insights.!! @rinnqd
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "620768": "Congrats to all the winners! Actually I expected to have bigger shakeup in the private LB but it seems that trusting the public LB was a good idea. I selected my top 2 LB and it worked : \n- 0.82 LB and in the private it gave 0.92 kappa \n- 0.817 LB and 0.92 in the private too\nThese both kernels are an ensemble of 4 models : \n- EfficientNet b0 , size=416 (btw this single model gave 0.806 in LB and 0.92 in the private)\n- EfficientNet b3 , size=288 (0.805 LB vs 0.913 private)\n- EfficientNet b4 , size=288 (did not submit it)\n- EfficientNet b5 , size=256 (0.797 LB vs 0.914 private)\nThe only difference between my top 2 kernels was the TTA:\nI just used a simple 5-TTA with these augmentations : \n ```\nalbumentations.Compose([\n albumentations.Resize(S, S),\n albumentations.ShiftScaleRotate(shift_limit=0.08, scale_limit=0.15, rotate_limit=30, p=0.5),\n albumentations.Normalize(),\n AT.ToTensor(),\n ])\n```\n\nbut gave me worse public LB. I thought TTA would help in the private LB but it seems that in all my TTA kernels scores went down even in private LB.\nAnother thing to add, all my models were trained on 80% of the data. I did not use Kfolds since Kfolds gave me bad results and I think I have chosen the golden fold that gave me more 0.02 score than the other folds. That's why ensembling did'nt gave me a huge boost since I am only validating my models on ~700 images.\nThank you for everyone who shared a kernel or an idea or made a comment or a post in the discussion section. \nThank you Kaggle and APTOS for this amazing competition.\n\n[Update : I trained every model on the old data of the previous competition (train+test) using the current data as validation. All my models converge to 0.88 kappa with pretrained weights]", "620910": "Congratulations\nGreat work\nAnd Thanks for sharing your approach and insights.!! @rinnqd" }, "source": "meta" }