{ "id": 201567, "title": "Image Augmentation or any sprinkle of magic?", "url": "/competitions/cassava-leaf-disease-classification/discussion/201567", "author_name": "Kamal Das", "post_date": "2020-12-05T16:03:57.947000", "votes": 6, "comment_count": 24, "views": 0, "content": "

Hi All
\nwhich Image Augmentation are proving helpful and adding to the score?
\nAnything new that you have tried and that has worked or not worked

\n

looking what what features or details has been able to help with the sprinkle of magic for your models…

\n

Thanks for sharing!

", "messages": [ { "id": 1103702, "postDate": "2020-12-06T06:39:31.357Z", "content": "

For me, good image augmentation really helped, but one thing that really boosted my CV is Progressive resizing technique.

", "rawMarkdown": "For me, good image augmentation really helped, but one thing that really boosted my CV is Progressive resizing technique.", "votes": 5, "replies": [ { "id": 1105963, "postDate": "2020-12-08T11:24:59.693Z", "content": "

Any pointer to literature/implementation of this technique?

", "rawMarkdown": "Any pointer to literature/implementation of this technique? ", "votes": 1 }, { "id": 1106181, "postDate": "2020-12-08T15:20:41.697Z", "content": "

I didnt find anything specific but: https://www.eurasip.org/Proceedings/Eusipco/Eusipco2020/pdfs/0000026.pdf trying to explain the idea behind it

\n

And \"code-wise\": https://github.com/ResidentMario/progressive-resizing

", "rawMarkdown": "I didnt find anything specific but: https://www.eurasip.org/Proceedings/Eusipco/Eusipco2020/pdfs/0000026.pdf trying to explain the idea behind it\n\nAnd \"code-wise\": https://github.com/ResidentMario/progressive-resizing", "votes": 1 }, { "id": 1106224, "postDate": "2020-12-08T16:02:37.680Z", "content": "

Noice, thank you mate, I will take a look.

", "rawMarkdown": "Noice, thank you mate, I will take a look.", "votes": 1 } ] }, { "id": 1103078, "postDate": "2020-12-05T16:03:57.947Z", "content": "

Hi All
\nwhich Image Augmentation are proving helpful and adding to the score?
\nAnything new that you have tried and that has worked or not worked

\n

looking what what features or details has been able to help with the sprinkle of magic for your models…

\n

Thanks for sharing!

", "rawMarkdown": "Hi All\nwhich Image Augmentation are proving helpful and adding to the score?\nAnything new that you have tried and that has worked or not worked\n\nlooking what what features or details has been able to help with the sprinkle of magic for your models...\n\nThanks for sharing!\n", "votes": 6 }, { "id": 1103669, "postDate": "2020-12-06T05:42:30.573Z", "content": "

Hi everyone, I don't know what STD and Mean you are using but if you are using the one mentioned in the baseline then try to change it according to your dataset that could increase your LB and CV.
\nIf you are using multiple models then taking the average of probabilities it is a good voting system but maybe some other ensemble classifiers could do better.

", "rawMarkdown": "Hi everyone, I don't know what STD and Mean you are using but if you are using the one mentioned in the baseline then try to change it according to your dataset that could increase your LB and CV.\nIf you are using multiple models then taking the average of probabilities it is a good voting system but maybe some other ensemble classifiers could do better.", "votes": 3, "replies": [ { "id": 1105724, "postDate": "2020-12-08T06:17:15.027Z", "content": "

why? I guess lots of people are using STD and Mean of ImageNet.

", "rawMarkdown": "why? I guess lots of people are using STD and Mean of ImageNet.", "votes": 1 }, { "id": 1105789, "postDate": "2020-12-08T07:43:36.440Z", "content": "

Feature scaling can screw things up, especially if some features are categorical. Anyway, I am not telling you to not use these values but it worth trying the std and mean for the competition dataset because the competition is not over ImageNet.

\n

mean: [0.43159878 0.49693648 0.31351838] , std: [0.23060184 0.23337172 0.22065296] ///256X256
\nmean: [0.42984136 0.49624753 0.3129598 ] , std: [0.23297946 0.2358761 0.22365381] ///512X512
\nAlso try without and choose the best CV.

", "rawMarkdown": "Feature scaling can screw things up, especially if some features are categorical. Anyway, I am not telling you to not use these values but it worth trying the std and mean for the competition dataset because the competition is not over ImageNet.\n\nmean: [0.43159878 0.49693648 0.31351838] , std: [0.23060184 0.23337172 0.22065296] ///256X256\nmean: [0.42984136 0.49624753 0.3129598 ] , std: [0.23297946 0.2358761 0.22365381] ///512X512\nAlso try without and choose the best CV.", "votes": 2 }, { "id": 1105844, "postDate": "2020-12-08T08:54:18.980Z", "content": "

Are these the mean and std of your dataset v2 ?

", "rawMarkdown": "Are these the mean and std of your dataset v2 ?", "votes": 2 }, { "id": 1105927, "postDate": "2020-12-08T10:39:24.257Z", "content": "

No these are for the original one.
\nhere are the values for mine:
\nmean: [0.46063745 0.54488302 0.33173344] , std: [0.22686065 0.22901212 0.22107227]

", "rawMarkdown": "No these are for the original one.\nhere are the values for mine:\nmean: [0.46063745 0.54488302 0.33173344] , std: [0.22686065 0.22901212 0.22107227]", "votes": 2 }, { "id": 1105929, "postDate": "2020-12-08T10:45:04.853Z", "content": "

Okay, I got mean=(0.46064209,0.5448876,0.33173755), std=(0.19857255,0.20062191,0.18294894)

\n

For 256x256 for your dataset

", "rawMarkdown": "Okay, I got mean=(0.46064209,0.5448876,0.33173755), std=(0.19857255,0.20062191,0.18294894)\n\nFor 256x256 for your dataset", "votes": 1 }, { "id": 1106031, "postDate": "2020-12-08T12:25:20.527Z", "content": "

thanks for pointing this out, will try it on CV and LB. Considered your rank, you must have a lot of other beautiful recipes, haha

", "rawMarkdown": "thanks for pointing this out, will try it on CV and LB. Considered your rank, you must have a lot of other beautiful recipes, haha", "votes": 1 } ] }, { "id": 1103614, "postDate": "2020-12-06T04:36:51.050Z", "content": "

Given your current LB score perhaps you can share what worked for you too :) For me CoarseDropout also worked.

", "rawMarkdown": "Given your current LB score perhaps you can share what worked for you too :) For me CoarseDropout also worked.", "votes": 4, "replies": [ { "id": 1103625, "postDate": "2020-12-06T04:46:20.983Z", "content": "

Actually, the image augmentation have not worked for me… no significant changes so was curious if some worked for you all
\nI got lucky by more training and also utilizing the extra dataset created by @ammarali32 here:
\nhttps://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200722

", "rawMarkdown": "Actually, the image augmentation have not worked for me... no significant changes so was curious if some worked for you all\nI got lucky by more training and also utilizing the extra dataset created by @ammarali32 here:\nhttps://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200722", "votes": 2 }, { "id": 1103636, "postDate": "2020-12-06T04:55:01.733Z", "content": "

Interesting, using that dataset might have accounted for some sort of data augmentation itself. But in general heavy augmentation helped me, not tried training longer epochs though, so my models still might be underfitting.

", "rawMarkdown": "Interesting, using that dataset might have accounted for some sort of data augmentation itself. But in general heavy augmentation helped me, not tried training longer epochs though, so my models still might be underfitting.", "votes": 1 }, { "id": 1103646, "postDate": "2020-12-06T05:04:27.530Z", "content": "

Hi.. sorry I use a standard set of augmentations like flips, random crops etc..
\nBut additional ones like affine etc did not help much beyond my original standard ones..
\nwhich augmentations did you find that added to your CV?

", "rawMarkdown": "Hi.. sorry I use a standard set of augmentations like flips, random crops etc..\nBut additional ones like affine etc did not help much beyond my original standard ones..\nwhich augmentations did you find that added to your CV?" }, { "id": 1103659, "postDate": "2020-12-06T05:21:39.180Z", "content": "

I also use flip, britghness, crop, .. regular transforms, I am now checking mixup and cutmix.

", "rawMarkdown": "I also use flip, britghness, crop, .. regular transforms, I am now checking mixup and cutmix.", "votes": 1 }, { "id": 1104231, "postDate": "2020-12-06T18:15:06.570Z", "content": "

I read people reporting that CutMix helped very well.

", "rawMarkdown": "I read people reporting that CutMix helped very well.", "votes": 1 } ] }, { "id": 1105291, "postDate": "2020-12-07T18:25:11.303Z", "content": "

I find cutmix relevant to this competition as it naturally creates new images of plants. Mixing plants with plants produces new plants :D

", "rawMarkdown": "I find cutmix relevant to this competition as it naturally creates new images of plants. Mixing plants with plants produces new plants :D", "votes": 1, "replies": [ { "id": 1105303, "postDate": "2020-12-07T18:41:51.877Z", "content": "

I wrote my own cutmix function which works good but the results are not better than my baseline

", "rawMarkdown": "I wrote my own cutmix function which works good but the results are not better than my baseline", "votes": 1 }, { "id": 1106274, "postDate": "2020-12-08T17:11:04.037Z", "content": "

this impl works best for me https://github.com/ildoonet/cutmix

", "rawMarkdown": "this impl works best for me https://github.com/ildoonet/cutmix", "votes": 1 }, { "id": 1106287, "postDate": "2020-12-08T17:18:29.133Z", "content": "

Looks equal to my function, do you use the CutMixCrossEntropyLoss? It looks interesting.

", "rawMarkdown": "Looks equal to my function, do you use the CutMixCrossEntropyLoss? It looks interesting.", "votes": 1 }, { "id": 1106294, "postDate": "2020-12-08T17:23:33.877Z", "content": "

yea. they state it works better with soft labels

", "rawMarkdown": "yea. they state it works better with soft labels", "votes": 1 } ] }, { "id": 1103179, "postDate": "2020-12-05T17:44:58.913Z", "content": "

What worked well for me is:

\n\n

I'm trying heavy augmentations but the results are not as good as expected.

", "rawMarkdown": "What worked well for me is:\n- CoarseDropout\n\nI'm trying heavy augmentations but the results are not as good as expected.\n\n", "votes": 2, "replies": [ { "id": 1103206, "postDate": "2020-12-05T18:04:07.947Z", "rawMarkdown": "", "isDeleted": true } ] } ], "comments": [ { "id": 1103702, "author_name": "DeepUnderstanding", "author_url": "", "post_date": "2020-12-06T06:39:31.357000", "content": "

For me, good image augmentation really helped, but one thing that really boosted my CV is Progressive resizing technique.

", "votes": 5, "replies": [ { "id": 1105963, "author_name": "Claudio Verdú Ruiz", "author_url": "", "post_date": "2020-12-08T11:24:59.693000", "content": "

Any pointer to literature/implementation of this technique?

", "votes": 1, "replies": [] }, { "id": 1106181, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-08T15:20:41.697000", "content": "

I didnt find anything specific but: https://www.eurasip.org/Proceedings/Eusipco/Eusipco2020/pdfs/0000026.pdf trying to explain the idea behind it

\n

And \"code-wise\": https://github.com/ResidentMario/progressive-resizing

", "votes": 1, "replies": [] }, { "id": 1106224, "author_name": "Claudio Verdú Ruiz", "author_url": "", "post_date": "2020-12-08T16:02:37.680000", "content": "

Noice, thank you mate, I will take a look.

", "votes": 1, "replies": [] } ] }, { "id": 1103669, "author_name": "ammarali32", "author_url": "", "post_date": "2020-12-06T05:42:30.573000", "content": "

Hi everyone, I don't know what STD and Mean you are using but if you are using the one mentioned in the baseline then try to change it according to your dataset that could increase your LB and CV.
\nIf you are using multiple models then taking the average of probabilities it is a good voting system but maybe some other ensemble classifiers could do better.

", "votes": 3, "replies": [ { "id": 1105724, "author_name": "Yi Wu", "author_url": "", "post_date": "2020-12-08T06:17:15.027000", "content": "

why? I guess lots of people are using STD and Mean of ImageNet.

", "votes": 1, "replies": [] }, { "id": 1105789, "author_name": "ammarali32", "author_url": "", "post_date": "2020-12-08T07:43:36.440000", "content": "

Feature scaling can screw things up, especially if some features are categorical. Anyway, I am not telling you to not use these values but it worth trying the std and mean for the competition dataset because the competition is not over ImageNet.

\n

mean: [0.43159878 0.49693648 0.31351838] , std: [0.23060184 0.23337172 0.22065296] ///256X256
\nmean: [0.42984136 0.49624753 0.3129598 ] , std: [0.23297946 0.2358761 0.22365381] ///512X512
\nAlso try without and choose the best CV.

", "votes": 2, "replies": [] }, { "id": 1105844, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-08T08:54:18.980000", "content": "

Are these the mean and std of your dataset v2 ?

", "votes": 2, "replies": [] }, { "id": 1105927, "author_name": "ammarali32", "author_url": "", "post_date": "2020-12-08T10:39:24.257000", "content": "

No these are for the original one.
\nhere are the values for mine:
\nmean: [0.46063745 0.54488302 0.33173344] , std: [0.22686065 0.22901212 0.22107227]

", "votes": 2, "replies": [] }, { "id": 1105929, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-08T10:45:04.853000", "content": "

Okay, I got mean=(0.46064209,0.5448876,0.33173755), std=(0.19857255,0.20062191,0.18294894)

\n

For 256x256 for your dataset

", "votes": 1, "replies": [] }, { "id": 1106031, "author_name": "Yi Wu", "author_url": "", "post_date": "2020-12-08T12:25:20.527000", "content": "

thanks for pointing this out, will try it on CV and LB. Considered your rank, you must have a lot of other beautiful recipes, haha

", "votes": 1, "replies": [] } ] }, { "id": 1103614, "author_name": "Kerem Turgutlu", "author_url": "", "post_date": "2020-12-06T04:36:51.050000", "content": "

Given your current LB score perhaps you can share what worked for you too :) For me CoarseDropout also worked.

", "votes": 4, "replies": [ { "id": 1103625, "author_name": "Kamal Das", "author_url": "", "post_date": "2020-12-06T04:46:20.983000", "content": "

Actually, the image augmentation have not worked for me… no significant changes so was curious if some worked for you all
\nI got lucky by more training and also utilizing the extra dataset created by @ammarali32 here:
\nhttps://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/200722

", "votes": 2, "replies": [] }, { "id": 1103636, "author_name": "Kerem Turgutlu", "author_url": "", "post_date": "2020-12-06T04:55:01.733000", "content": "

Interesting, using that dataset might have accounted for some sort of data augmentation itself. But in general heavy augmentation helped me, not tried training longer epochs though, so my models still might be underfitting.

", "votes": 1, "replies": [] }, { "id": 1103646, "author_name": "Kamal Das", "author_url": "", "post_date": "2020-12-06T05:04:27.530000", "content": "

Hi.. sorry I use a standard set of augmentations like flips, random crops etc..
\nBut additional ones like affine etc did not help much beyond my original standard ones..
\nwhich augmentations did you find that added to your CV?

", "votes": 0, "replies": [] }, { "id": 1103659, "author_name": "Kerem Turgutlu", "author_url": "", "post_date": "2020-12-06T05:21:39.180000", "content": "

I also use flip, britghness, crop, .. regular transforms, I am now checking mixup and cutmix.

", "votes": 1, "replies": [] }, { "id": 1104231, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-06T18:15:06.570000", "content": "

I read people reporting that CutMix helped very well.

", "votes": 1, "replies": [] } ] }, { "id": 1105291, "author_name": "Georgiy Grigorev", "author_url": "", "post_date": "2020-12-07T18:25:11.303000", "content": "

I find cutmix relevant to this competition as it naturally creates new images of plants. Mixing plants with plants produces new plants :D

", "votes": 1, "replies": [ { "id": 1105303, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-07T18:41:51.877000", "content": "

I wrote my own cutmix function which works good but the results are not better than my baseline

", "votes": 1, "replies": [] }, { "id": 1106274, "author_name": "Georgiy Grigorev", "author_url": "", "post_date": "2020-12-08T17:11:04.037000", "content": "

this impl works best for me https://github.com/ildoonet/cutmix

", "votes": 1, "replies": [] }, { "id": 1106287, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-08T17:18:29.133000", "content": "

Looks equal to my function, do you use the CutMixCrossEntropyLoss? It looks interesting.

", "votes": 1, "replies": [] }, { "id": 1106294, "author_name": "Georgiy Grigorev", "author_url": "", "post_date": "2020-12-08T17:23:33.877000", "content": "

yea. they state it works better with soft labels

", "votes": 1, "replies": [] } ] }, { "id": 1103179, "author_name": "Ali Abdin", "author_url": "", "post_date": "2020-12-05T17:44:58.913000", "content": "

What worked well for me is:

\n\n

I'm trying heavy augmentations but the results are not as good as expected.

", "votes": 2, "replies": [ { "id": 1103206, "author_name": "", "author_url": "", "post_date": "2020-12-05T18:04:07.947000", "content": "", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1103702": "For me, good image augmentation really helped, but one thing that really boosted my CV is Progressive resizing technique.", "1103078": "Hi All\nwhich Image Augmentation are proving helpful and adding to the score?\nAnything new that you have tried and that has worked or not worked\n\nlooking what what features or details has been able to help with the sprinkle of magic for your models...\n\nThanks for sharing!\n", "1103669": "Hi everyone, I don't know what STD and Mean you are using but if you are using the one mentioned in the baseline then try to change it according to your dataset that could increase your LB and CV.\nIf you are using multiple models then taking the average of probabilities it is a good voting system but maybe some other ensemble classifiers could do better.", "1103614": "Given your current LB score perhaps you can share what worked for you too :) For me CoarseDropout also worked.", "1105291": "I find cutmix relevant to this competition as it naturally creates new images of plants. Mixing plants with plants produces new plants :D", "1103179": "What worked well for me is:\n- CoarseDropout\n\nI'm trying heavy augmentations but the results are not as good as expected.\n\n" } }