{ "id": 202946, "title": "Getting LB score of 0.048 even after using everything to the best of my knowledge.", "url": "/competitions/cassava-leaf-disease-classification/discussion/202946", "author_name": "", "post_date": "2020-12-12T20:36:30.662190900Z", "votes": null, "comment_count": 6, "views": 0, "content": "

Hello All,

\n

I am unable to understand why I getting such a low score on LB of 0.048 even after using these:

\n

--> using image size of 256 * 256
\n--> following train and validation augmentations with FMix

\n
if self.is_valid == 1: # transforms for validation images\n            self.aug = albumentations.Compose([\n            albumentations.CenterCrop(256, 256, p=1.),\n            albumentations.Resize(256, 256),\n            albumentations.Normalize(\n                mean=[0.485, 0.456, 0.406], \n                std=[0.229, 0.224, 0.225], \n                max_pixel_value=255.0, \n                p=1.0\n            )], p=1.)\n\n        else:       # transfoms for training images \n            self.aug = albumentations.Compose([\n            albumentations.RandomResizedCrop(256, 256),\n            albumentations.Transpose(p=0.5),\n            albumentations.HorizontalFlip(p=0.5),\n            albumentations.VerticalFlip(p=0.5),\n            albumentations.ShiftScaleRotate(p=0.5),\n            albumentations.HueSaturationValue(\n                hue_shift_limit=0.2, \n                sat_shift_limit=0.2, \n                val_shift_limit=0.2, \n                p=0.5\n            ),\n            albumentations.RandomBrightnessContrast(\n                brightness_limit=(-0.1,0.1), \n                contrast_limit=(-0.1, 0.1), \n                p=0.5\n            ),\n            albumentations.Normalize(\n                mean=[0.485, 0.456, 0.406], \n                std=[0.229, 0.224, 0.225], \n                max_pixel_value=255.0, \n                p=1.0\n            ),\n            albumentations.CoarseDropout(p=0.5),\n            albumentations.Cutout(p=0.5)], p=1.)\n
\n

--> EfficientNet-b4 with addition of one more extra layer of adaptive average pooling
\n--> Learning rate of 1e-3 * 0.95
\n--> Loss function as Cross Entropy Loss
\n--> Scheduler as ReduceLROnPlateau
\n--> Stochastic Weight averaging
\n--> Finally Test Time Augmentation while doing inference

\n

IN PyTorch getting a score of 82% max on the given dataset
\nMy kernel Link

\n

Can anyone please help me out in telling where I am doing wrong?
\nI will be highly thankful

", "messages": [ { "id": "1110529", "postDate": "12/12/2020 20:36:30", "content": "

Hello All,

\n

I am unable to understand why I getting such a low score on LB of 0.048 even after using these:

\n

--> using image size of 256 * 256
\n--> following train and validation augmentations with FMix

\n
if self.is_valid == 1: # transforms for validation images\n            self.aug = albumentations.Compose([\n            albumentations.CenterCrop(256, 256, p=1.),\n            albumentations.Resize(256, 256),\n            albumentations.Normalize(\n                mean=[0.485, 0.456, 0.406], \n                std=[0.229, 0.224, 0.225], \n                max_pixel_value=255.0, \n                p=1.0\n            )], p=1.)\n\n        else:       # transfoms for training images \n            self.aug = albumentations.Compose([\n            albumentations.RandomResizedCrop(256, 256),\n            albumentations.Transpose(p=0.5),\n            albumentations.HorizontalFlip(p=0.5),\n            albumentations.VerticalFlip(p=0.5),\n            albumentations.ShiftScaleRotate(p=0.5),\n            albumentations.HueSaturationValue(\n                hue_shift_limit=0.2, \n                sat_shift_limit=0.2, \n                val_shift_limit=0.2, \n                p=0.5\n            ),\n            albumentations.RandomBrightnessContrast(\n                brightness_limit=(-0.1,0.1), \n                contrast_limit=(-0.1, 0.1), \n                p=0.5\n            ),\n            albumentations.Normalize(\n                mean=[0.485, 0.456, 0.406], \n                std=[0.229, 0.224, 0.225], \n                max_pixel_value=255.0, \n                p=1.0\n            ),\n            albumentations.CoarseDropout(p=0.5),\n            albumentations.Cutout(p=0.5)], p=1.)\n
\n

--> EfficientNet-b4 with addition of one more extra layer of adaptive average pooling
\n--> Learning rate of 1e-3 * 0.95
\n--> Loss function as Cross Entropy Loss
\n--> Scheduler as ReduceLROnPlateau
\n--> Stochastic Weight averaging
\n--> Finally Test Time Augmentation while doing inference

\n

IN PyTorch getting a score of 82% max on the given dataset
\nMy kernel Link

\n

Can anyone please help me out in telling where I am doing wrong?
\nI will be highly thankful

", "rawMarkdown": "Hello All,\n\nI am unable to understand why I getting such a low score on LB of 0.048 even after using these:\n\n--> using image size of 256 * 256\n--> following train and validation augmentations with FMix\n```\nif self.is_valid == 1: # transforms for validation images\n self.aug = albumentations.Compose([\n albumentations.CenterCrop(256, 256, p=1.),\n albumentations.Resize(256, 256),\n albumentations.Normalize(\n mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225], \n max_pixel_value=255.0, \n p=1.0\n )], p=1.)\n\n else: # transfoms for training images \n self.aug = albumentations.Compose([\n albumentations.RandomResizedCrop(256, 256),\n albumentations.Transpose(p=0.5),\n albumentations.HorizontalFlip(p=0.5),\n albumentations.VerticalFlip(p=0.5),\n albumentations.ShiftScaleRotate(p=0.5),\n albumentations.HueSaturationValue(\n hue_shift_limit=0.2, \n sat_shift_limit=0.2, \n val_shift_limit=0.2, \n p=0.5\n ),\n albumentations.RandomBrightnessContrast(\n brightness_limit=(-0.1,0.1), \n contrast_limit=(-0.1, 0.1), \n p=0.5\n ),\n albumentations.Normalize(\n mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225], \n max_pixel_value=255.0, \n p=1.0\n ),\n albumentations.CoarseDropout(p=0.5),\n albumentations.Cutout(p=0.5)], p=1.)\n```\n--> EfficientNet-b4 with addition of one more extra layer of adaptive average pooling\n--> Learning rate of 1e-3 * 0.95\n--> Loss function as Cross Entropy Loss\n--> Scheduler as ReduceLROnPlateau\n--> Stochastic Weight averaging\n--> Finally Test Time Augmentation while doing inference \n\nIN PyTorch getting a score of 82% max on the given dataset\n[My kernel Link ](https://www.kaggle.com/soumochatterjee/cassava-gpu-fmix-efficientnet-b4)\n\nCan anyone please help me out in telling where I am doing wrong?\nI will be highly thankful", "votes": null }, { "id": "1110700", "postDate": "12/13/2020 01:13:59", "content": "

Hello!
\nIs 256 * 256 better than 512 * 512 ?

", "rawMarkdown": "Hello!\nIs 256 * 256 better than 512 * 512 ?", "votes": null }, { "id": "1110925", "postDate": "12/13/2020 07:52:13", "content": "

Do check your images after doing all this augmentations and see if it is not giving completely unrecognisable image;
\nalso change your Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) its different in this dataset

", "rawMarkdown": "Do check your images after doing all this augmentations and see if it is not giving completely unrecognisable image;\nalso change your Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) its different in this dataset", "votes": null }, { "id": "1110929", "postDate": "12/13/2020 08:01:41", "content": "

@mrinath I have posted my kernel here, please have a look to it . I did check my images after setting the custom dataset and what value of mean and standard deviation should I use? Can u please mention it !

", "rawMarkdown": "mrinath I have posted my kernel here, please have a look to it . I did check my images after setting the custom dataset and what value of mean and standard deviation should I use? Can u please mention it !", "votes": null }, { "id": "1111078", "postDate": "12/13/2020 11:45:10", "content": "

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199980
\nrefer to this

", "rawMarkdown": "https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199980\nrefer to this", "votes": null }, { "id": "1111149", "postDate": "12/13/2020 13:11:24", "content": "

Only in so far as you can process the smaller size faster (and perhaps can make do with a less deep model), beyond that if computational resources is not the issue, bigger is better. The efficientnet paper discusses the trade off and ideas for ideal scaling (also for other architectures such as resent).

", "rawMarkdown": "Only in so far as you can process the smaller size faster (and perhaps can make do with a less deep model), beyond that if computational resources is not the issue, bigger is better. The [efficientnet paper](https://arxiv.org/abs/1905.11946) discusses the trade off and ideas for ideal scaling (also for other architectures such as resent).", "votes": null }, { "id": "1111302", "postDate": "12/13/2020 16:06:06", "content": "

Your inference is putting all prediction as 0. There is a bug in your inference (prediction) logic. Check below thread for the testing dataset profile:
\n https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202943

", "rawMarkdown": "Your inference is putting all prediction as 0. There is a bug in your inference (prediction) logic. Check below thread for the testing dataset profile:\n https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202943", "votes": null } ], "comments": [ { "id": 1110700, "author_name": "zekunn", "author_url": "", "post_date": "12/13/2020 01:13:59", "content": "

Hello!
\nIs 256 * 256 better than 512 * 512 ?

", "votes": null, "replies": [ { "id": 1111149, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "12/13/2020 13:11:24", "content": "

Only in so far as you can process the smaller size faster (and perhaps can make do with a less deep model), beyond that if computational resources is not the issue, bigger is better. The efficientnet paper discusses the trade off and ideas for ideal scaling (also for other architectures such as resent).

", "votes": null, "replies": [] } ] }, { "id": 1110925, "author_name": "mrinath", "author_url": "", "post_date": "12/13/2020 07:52:13", "content": "

Do check your images after doing all this augmentations and see if it is not giving completely unrecognisable image;
\nalso change your Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) its different in this dataset

", "votes": null, "replies": [] }, { "id": 1110929, "author_name": "soumochatterjee", "author_url": "", "post_date": "12/13/2020 08:01:41", "content": "

@mrinath I have posted my kernel here, please have a look to it . I did check my images after setting the custom dataset and what value of mean and standard deviation should I use? Can u please mention it !

", "votes": null, "replies": [ { "id": 1111078, "author_name": "mrinath", "author_url": "", "post_date": "12/13/2020 11:45:10", "content": "

https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199980
\nrefer to this

", "votes": null, "replies": [] } ] }, { "id": 1111302, "author_name": "luqing2", "author_url": "", "post_date": "12/13/2020 16:06:06", "content": "

Your inference is putting all prediction as 0. There is a bug in your inference (prediction) logic. Check below thread for the testing dataset profile:
\n https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202943

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1110529": "Hello All,\n\nI am unable to understand why I getting such a low score on LB of 0.048 even after using these:\n\n--> using image size of 256 * 256\n--> following train and validation augmentations with FMix\n```\nif self.is_valid == 1: # transforms for validation images\n self.aug = albumentations.Compose([\n albumentations.CenterCrop(256, 256, p=1.),\n albumentations.Resize(256, 256),\n albumentations.Normalize(\n mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225], \n max_pixel_value=255.0, \n p=1.0\n )], p=1.)\n\n else: # transfoms for training images \n self.aug = albumentations.Compose([\n albumentations.RandomResizedCrop(256, 256),\n albumentations.Transpose(p=0.5),\n albumentations.HorizontalFlip(p=0.5),\n albumentations.VerticalFlip(p=0.5),\n albumentations.ShiftScaleRotate(p=0.5),\n albumentations.HueSaturationValue(\n hue_shift_limit=0.2, \n sat_shift_limit=0.2, \n val_shift_limit=0.2, \n p=0.5\n ),\n albumentations.RandomBrightnessContrast(\n brightness_limit=(-0.1,0.1), \n contrast_limit=(-0.1, 0.1), \n p=0.5\n ),\n albumentations.Normalize(\n mean=[0.485, 0.456, 0.406], \n std=[0.229, 0.224, 0.225], \n max_pixel_value=255.0, \n p=1.0\n ),\n albumentations.CoarseDropout(p=0.5),\n albumentations.Cutout(p=0.5)], p=1.)\n```\n--> EfficientNet-b4 with addition of one more extra layer of adaptive average pooling\n--> Learning rate of 1e-3 * 0.95\n--> Loss function as Cross Entropy Loss\n--> Scheduler as ReduceLROnPlateau\n--> Stochastic Weight averaging\n--> Finally Test Time Augmentation while doing inference \n\nIN PyTorch getting a score of 82% max on the given dataset\n[My kernel Link ](https://www.kaggle.com/soumochatterjee/cassava-gpu-fmix-efficientnet-b4)\n\nCan anyone please help me out in telling where I am doing wrong?\nI will be highly thankful", "1110700": "Hello!\nIs 256 * 256 better than 512 * 512 ?", "1110925": "Do check your images after doing all this augmentations and see if it is not giving completely unrecognisable image;\nalso change your Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) its different in this dataset", "1110929": "mrinath I have posted my kernel here, please have a look to it . I did check my images after setting the custom dataset and what value of mean and standard deviation should I use? Can u please mention it !", "1111078": "https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/199980\nrefer to this", "1111149": "Only in so far as you can process the smaller size faster (and perhaps can make do with a less deep model), beyond that if computational resources is not the issue, bigger is better. The [efficientnet paper](https://arxiv.org/abs/1905.11946) discusses the trade off and ideas for ideal scaling (also for other architectures such as resent).", "1111302": "Your inference is putting all prediction as 0. There is a bug in your inference (prediction) logic. Check below thread for the testing dataset profile:\n https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202943" }, "source": "meta" }