{ "id": 206489, "title": "Best TTA for me is almost no aug", "url": "/competitions/cassava-leaf-disease-classification/discussion/206489", "author_name": "", "post_date": "2020-12-24T22:46:07.995715600Z", "votes": 37, "comment_count": 17, "views": 0, "content": "

It is quite usual to see heavy augmentations for the Robustness in CV task, and even heavy TTA on Kaggle for a higher place on LB.

\n

However, after 20+ submissions this week, I noticed that light (or you can say almost none) inference augmentation (just resize&crop and normalize) socred the best which surprised me a lot.

\n

Commonly used methods like flip, transpose, brightness etc. did not give me a stable improvement. (actually somehow decreased the performance for me)

\n

Wondering what happened on your guys.

\n

We'd better carefully select augmentations while doing TTA, or even reconsider augmentations while training.

", "messages": [ { "id": "1125643", "postDate": "12/24/2020 22:46:07", "content": "

It is quite usual to see heavy augmentations for the Robustness in CV task, and even heavy TTA on Kaggle for a higher place on LB.

\n

However, after 20+ submissions this week, I noticed that light (or you can say almost none) inference augmentation (just resize&crop and normalize) socred the best which surprised me a lot.

\n

Commonly used methods like flip, transpose, brightness etc. did not give me a stable improvement. (actually somehow decreased the performance for me)

\n

Wondering what happened on your guys.

\n

We'd better carefully select augmentations while doing TTA, or even reconsider augmentations while training.

", "rawMarkdown": "It is quite usual to see heavy augmentations for the Robustness in CV task, and even heavy TTA on Kaggle for a higher place on LB.\n\nHowever, after 20+ submissions this week, I noticed that light (or you can say almost none) inference augmentation (just resize&crop and normalize) socred the best which surprised me a lot.\n\nCommonly used methods like flip, transpose, brightness etc. did not give me a stable improvement. (actually somehow decreased the performance for me)\n\nWondering what happened on your guys.\n\nWe'd better carefully select augmentations while doing TTA, or even reconsider augmentations while training.", "votes": null }, { "id": "1125685", "postDate": "12/25/2020 00:43:14", "content": "

I have the same feelings.

", "rawMarkdown": "I have the same feelings.", "votes": null }, { "id": "1125707", "postDate": "12/25/2020 01:37:01", "content": "

i also don't work flip, tanspose

", "rawMarkdown": "i also don't work flip, tanspose", "votes": null }, { "id": "1125914", "postDate": "12/25/2020 07:09:56", "content": "

In my case just transposing, flipping, and the usual resizing, normalizing increased my LB by 0.2;

", "rawMarkdown": "In my case just transposing, flipping, and the usual resizing, normalizing increased my LB by 0.2;", "votes": null }, { "id": "1125971", "postDate": "12/25/2020 07:58:31", "content": "

I think flipping, transposing and combinations can be helpful, others may be harmful, after all have risks losing image info

", "rawMarkdown": "I think flipping, transposing and combinations can be helpful, others may be harmful, after all have risks losing image info", "votes": null }, { "id": "1126127", "postDate": "12/25/2020 11:18:14", "content": "

i removed some augmentations after seeing this and give it a try and it boost my score. nice catch

", "rawMarkdown": "i removed some augmentations after seeing this and give it a try and it boost my score. nice catch", "votes": null }, { "id": "1126178", "postDate": "12/25/2020 12:09:14", "content": "

Thanks sharing experiment results!

", "rawMarkdown": "Thanks sharing experiment results!", "votes": null }, { "id": "1126826", "postDate": "12/26/2020 02:22:00", "content": "

Did this also happen in your CV? or just LB?

", "rawMarkdown": "Did this also happen in your CV? or just LB?", "votes": null }, { "id": "1126871", "postDate": "12/26/2020 04:07:49", "content": "

I did not use the same TTA while valid, not sure about that.
\nSo it is just LB.
\nThe same CV could score both 0.902 and 0.905 w/wo TTA.
\nHowever, considering the size of the public LB, the improvement was not a big deal but still worth notice.

", "rawMarkdown": "I did not use the same TTA while valid, not sure about that. \nSo it is just LB.\nThe same CV could score both 0.902 and 0.905 w/wo TTA.\nHowever, considering the size of the public LB, the improvement was not a big deal but still worth notice.", "votes": null }, { "id": "1126873", "postDate": "12/26/2020 04:10:17", "content": "

Hi, all the folds can achieve [0.902, 0.905]?

", "rawMarkdown": "Hi, all the folds can achieve [0.902, 0.905]?", "votes": null }, { "id": "1126955", "postDate": "12/26/2020 05:57:06", "content": "

Currently, I only test single fold but self-ensemble different epochs. It was a VIT + single fold effnet

", "rawMarkdown": "Currently, I only test single fold but self-ensemble different epochs. It was a VIT + single fold effnet", "votes": null }, { "id": "1127015", "postDate": "12/26/2020 07:06:51", "content": "

flip , transpose works fine with me @zlanan

", "rawMarkdown": "flip , transpose works fine with me @zlanan", "votes": null }, { "id": "1127265", "postDate": "12/26/2020 11:31:38", "content": "

For me heavy augmentations helped to improve my CV but not LB.

", "rawMarkdown": "For me heavy augmentations helped to improve my CV but not LB.", "votes": null }, { "id": "1127284", "postDate": "12/26/2020 11:59:27", "content": "

I had a better valid loss with more augs, but similar valid accuracy.

", "rawMarkdown": "I had a better valid loss with more augs, but similar valid accuracy.", "votes": null }, { "id": "1128645", "postDate": "12/27/2020 16:19:01", "content": "

While training did you use light augmentations too?

", "rawMarkdown": "While training did you use light augmentations too?", "votes": null }, { "id": "1128737", "postDate": "12/27/2020 17:25:41", "content": "

For me Hflip, VFlip, Transposed increased my CV by 0.0015

", "rawMarkdown": "For me Hflip, VFlip, Transposed increased my CV by 0.0015", "votes": null }, { "id": "1132229", "postDate": "12/30/2020 08:51:26", "content": "

Hi, what do you mean by 'self-ensemble different epochs' ? thx

", "rawMarkdown": "Hi, what do you mean by 'self-ensemble different epochs' ? thx", "votes": null }, { "id": "1132260", "postDate": "12/30/2020 09:17:02", "content": "

I think he means SWA

", "rawMarkdown": "I think he means SWA", "votes": null } ], "comments": [ { "id": 1125685, "author_name": "zekunn", "author_url": "", "post_date": "12/25/2020 00:43:14", "content": "

I have the same feelings.

", "votes": null, "replies": [] }, { "id": 1125707, "author_name": "chocozzz", "author_url": "", "post_date": "12/25/2020 01:37:01", "content": "

i also don't work flip, tanspose

", "votes": null, "replies": [] }, { "id": 1125914, "author_name": "mrinath", "author_url": "", "post_date": "12/25/2020 07:09:56", "content": "

In my case just transposing, flipping, and the usual resizing, normalizing increased my LB by 0.2;

", "votes": null, "replies": [] }, { "id": 1125971, "author_name": "plugin1689", "author_url": "", "post_date": "12/25/2020 07:58:31", "content": "

I think flipping, transposing and combinations can be helpful, others may be harmful, after all have risks losing image info

", "votes": null, "replies": [] }, { "id": 1126127, "author_name": "erntkn", "author_url": "", "post_date": "12/25/2020 11:18:14", "content": "

i removed some augmentations after seeing this and give it a try and it boost my score. nice catch

", "votes": null, "replies": [ { "id": 1126178, "author_name": "chocozzz", "author_url": "", "post_date": "12/25/2020 12:09:14", "content": "

Thanks sharing experiment results!

", "votes": null, "replies": [] } ] }, { "id": 1126826, "author_name": "yannmajewski", "author_url": "", "post_date": "12/26/2020 02:22:00", "content": "

Did this also happen in your CV? or just LB?

", "votes": null, "replies": [ { "id": 1126871, "author_name": "zlanan", "author_url": "", "post_date": "12/26/2020 04:07:49", "content": "

I did not use the same TTA while valid, not sure about that.
\nSo it is just LB.
\nThe same CV could score both 0.902 and 0.905 w/wo TTA.
\nHowever, considering the size of the public LB, the improvement was not a big deal but still worth notice.

", "votes": null, "replies": [] }, { "id": 1126873, "author_name": "plugin1689", "author_url": "", "post_date": "12/26/2020 04:10:17", "content": "

Hi, all the folds can achieve [0.902, 0.905]?

", "votes": null, "replies": [] }, { "id": 1126955, "author_name": "zlanan", "author_url": "", "post_date": "12/26/2020 05:57:06", "content": "

Currently, I only test single fold but self-ensemble different epochs. It was a VIT + single fold effnet

", "votes": null, "replies": [] }, { "id": 1132229, "author_name": "clwwlc", "author_url": "", "post_date": "12/30/2020 08:51:26", "content": "

Hi, what do you mean by 'self-ensemble different epochs' ? thx

", "votes": null, "replies": [] }, { "id": 1132260, "author_name": "plugin1689", "author_url": "", "post_date": "12/30/2020 09:17:02", "content": "

I think he means SWA

", "votes": null, "replies": [] } ] }, { "id": 1127015, "author_name": "saurabhshahane", "author_url": "", "post_date": "12/26/2020 07:06:51", "content": "

flip , transpose works fine with me @zlanan

", "votes": null, "replies": [] }, { "id": 1127265, "author_name": "nroman", "author_url": "", "post_date": "12/26/2020 11:31:38", "content": "

For me heavy augmentations helped to improve my CV but not LB.

", "votes": null, "replies": [ { "id": 1127284, "author_name": "zlanan", "author_url": "", "post_date": "12/26/2020 11:59:27", "content": "

I had a better valid loss with more augs, but similar valid accuracy.

", "votes": null, "replies": [] } ] }, { "id": 1128645, "author_name": "mrinath", "author_url": "", "post_date": "12/27/2020 16:19:01", "content": "

While training did you use light augmentations too?

", "votes": null, "replies": [] }, { "id": 1128737, "author_name": "yannmajewski", "author_url": "", "post_date": "12/27/2020 17:25:41", "content": "

For me Hflip, VFlip, Transposed increased my CV by 0.0015

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1125643": "It is quite usual to see heavy augmentations for the Robustness in CV task, and even heavy TTA on Kaggle for a higher place on LB.\n\nHowever, after 20+ submissions this week, I noticed that light (or you can say almost none) inference augmentation (just resize&crop and normalize) socred the best which surprised me a lot.\n\nCommonly used methods like flip, transpose, brightness etc. did not give me a stable improvement. (actually somehow decreased the performance for me)\n\nWondering what happened on your guys.\n\nWe'd better carefully select augmentations while doing TTA, or even reconsider augmentations while training.", "1125685": "I have the same feelings.", "1125707": "i also don't work flip, tanspose", "1125914": "In my case just transposing, flipping, and the usual resizing, normalizing increased my LB by 0.2;", "1125971": "I think flipping, transposing and combinations can be helpful, others may be harmful, after all have risks losing image info", "1126127": "i removed some augmentations after seeing this and give it a try and it boost my score. nice catch", "1126178": "Thanks sharing experiment results!", "1126826": "Did this also happen in your CV? or just LB?", "1126871": "I did not use the same TTA while valid, not sure about that. \nSo it is just LB.\nThe same CV could score both 0.902 and 0.905 w/wo TTA.\nHowever, considering the size of the public LB, the improvement was not a big deal but still worth notice.", "1126873": "Hi, all the folds can achieve [0.902, 0.905]?", "1126955": "Currently, I only test single fold but self-ensemble different epochs. It was a VIT + single fold effnet", "1127015": "flip , transpose works fine with me @zlanan", "1127265": "For me heavy augmentations helped to improve my CV but not LB.", "1127284": "I had a better valid loss with more augs, but similar valid accuracy.", "1128645": "While training did you use light augmentations too?", "1128737": "For me Hflip, VFlip, Transposed increased my CV by 0.0015", "1132229": "Hi, what do you mean by 'self-ensemble different epochs' ? thx", "1132260": "I think he means SWA" }, "source": "meta" }