{ "id": 428163, "title": "Be Careful If You Use TTA for Segmentation Task", "url": "/competitions/google-research-identify-contrails-reduce-global-warming/discussion/428163", "author_name": "william.wu", "post_date": "2023-07-31T11:20:04.637000", "votes": 9, "comment_count": 5, "views": 0, "content": "

Some augmentation like flipping and rotating alter the image orientation, thus also changing the orientation of the predictions.

\n

\"HFlip

\n

You need to deaug the prediction of aug Image before ensembling it with the prediction of the original one. In my pipeline, I do it in this way:

\n
hflip = A.HorizontalFlip(p=)\nvflip = A.VerticalFlip(p=)\naugs = [, hflip, vflip]\nde_augs = [, hflip, vflip]\nall_preds = []\n aug  augs:\n    all_preds.append(predict(model, data_loader, transform=aug)\n i, de_aug  (de_augs):\n    all_preds[i] = transform(all_preds[i], transform=de_aug)\npreds = np.average(all_preds, axis=)\n
", "messages": [ { "id": 2367158, "postDate": "2023-07-31T11:50:19.500Z", "content": "

just use https://github.com/BloodAxe/pytorch-toolbelt

", "rawMarkdown": "just use https://github.com/BloodAxe/pytorch-toolbelt", "votes": 10, "replies": [ { "id": 2367287, "postDate": "2023-07-31T13:25:54.417Z", "content": "

or https://github.com/qubvel/ttach

", "rawMarkdown": "or https://github.com/qubvel/ttach", "votes": 10, "replies": [ { "id": 2367319, "postDate": "2023-07-31T13:49:45.580Z", "content": "

Thanks for sharing, it's really helpful

", "rawMarkdown": "Thanks for sharing, it's really helpful", "votes": 1 }, { "id": 2381620, "postDate": "2023-08-09T09:40:03.553Z", "content": "

@maksimovka @sergiosaharovskiy Please guide me to use this useful repo on kaggle when internet off?

", "rawMarkdown": "@maksimovka @sergiosaharovskiy Please guide me to use this useful repo on kaggle when internet off?", "votes": 1 }, { "id": 2381639, "postDate": "2023-08-09T09:56:11.357Z", "content": "

just use a dataset which holds the library code
\nhttps://www.kaggle.com/datasets/sergiosaharovskiy/ttach

", "rawMarkdown": "just use a dataset which holds the library code\nhttps://www.kaggle.com/datasets/sergiosaharovskiy/ttach", "votes": 2 } ] } ] }, { "id": 2367127, "postDate": "2023-07-31T11:20:04.637Z", "content": "

Some augmentation like flipping and rotating alter the image orientation, thus also changing the orientation of the predictions.

\n

\"HFlip

\n

You need to deaug the prediction of aug Image before ensembling it with the prediction of the original one. In my pipeline, I do it in this way:

\n
hflip = A.HorizontalFlip(p=)\nvflip = A.VerticalFlip(p=)\naugs = [, hflip, vflip]\nde_augs = [, hflip, vflip]\nall_preds = []\n aug  augs:\n    all_preds.append(predict(model, data_loader, transform=aug)\n i, de_aug  (de_augs):\n    all_preds[i] = transform(all_preds[i], transform=de_aug)\npreds = np.average(all_preds, axis=)\n
", "rawMarkdown": "Some augmentation like flipping and rotating alter the image orientation, thus also changing the orientation of the predictions. \n\n![HFlip Aug](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3563032%2F300bf5fe48f935c50fc6403dcb1f5664%2Faug.png?generation=1690802059493745&alt=media)\n\nYou need to deaug the prediction of aug Image before ensembling it with the prediction of the original one. In my pipeline, I do it in this way:\n```Python\nhflip = A.HorizontalFlip(p=1)\nvflip = A.VerticalFlip(p=1)\naugs = [None, hflip, vflip]\nde_augs = [None, hflip, vflip]\nall_preds = []\nfor aug in augs:\n all_preds.append(predict(model, data_loader, transform=aug)\nfor i, de_aug in enumerate(de_augs):\n all_preds[i] = transform(all_preds[i], transform=de_aug)\npreds = np.average(all_preds, axis=0)\n```", "votes": 8 } ], "comments": [ { "id": 2367158, "author_name": "Kostiantyn Maksymov", "author_url": "", "post_date": "2023-07-31T11:50:19.500000", "content": "

just use https://github.com/BloodAxe/pytorch-toolbelt

", "votes": 10, "replies": [ { "id": 2367287, "author_name": "SSS", "author_url": "", "post_date": "2023-07-31T13:25:54.417000", "content": "

or https://github.com/qubvel/ttach

", "votes": 10, "replies": [ { "id": 2367319, "author_name": "william.wu", "author_url": "", "post_date": "2023-07-31T13:49:45.580000", "content": "

Thanks for sharing, it's really helpful

", "votes": 1, "replies": [] }, { "id": 2381620, "author_name": "Vipin Kumar", "author_url": "", "post_date": "2023-08-09T09:40:03.553000", "content": "

@maksimovka @sergiosaharovskiy Please guide me to use this useful repo on kaggle when internet off?

", "votes": 1, "replies": [] }, { "id": 2381639, "author_name": "iadduk", "author_url": "", "post_date": "2023-08-09T09:56:11.357000", "content": "

just use a dataset which holds the library code
\nhttps://www.kaggle.com/datasets/sergiosaharovskiy/ttach

", "votes": 2, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2367158": "just use https://github.com/BloodAxe/pytorch-toolbelt", "2367127": "Some augmentation like flipping and rotating alter the image orientation, thus also changing the orientation of the predictions. \n\n![HFlip Aug](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3563032%2F300bf5fe48f935c50fc6403dcb1f5664%2Faug.png?generation=1690802059493745&alt=media)\n\nYou need to deaug the prediction of aug Image before ensembling it with the prediction of the original one. In my pipeline, I do it in this way:\n```Python\nhflip = A.HorizontalFlip(p=1)\nvflip = A.VerticalFlip(p=1)\naugs = [None, hflip, vflip]\nde_augs = [None, hflip, vflip]\nall_preds = []\nfor aug in augs:\n all_preds.append(predict(model, data_loader, transform=aug)\nfor i, de_aug in enumerate(de_augs):\n all_preds[i] = transform(all_preds[i], transform=de_aug)\npreds = np.average(all_preds, axis=0)\n```" } }