{ "id": 348422, "title": "3D data augmentations library", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/348422", "author_name": "", "post_date": "2022-08-28T10:51:23.724801300Z", "votes": 4, "comment_count": 5, "views": 0, "content": "

Is anyone aware of a library that performs data augmentations on 3d volumes, compatible with PyTorch and ideally works on GPU?

\n

During training, I'm loading (B,C,D,H,W)=(batch_size,1,224,224,224) tensors and I'd like to apply rotations/flips in the x-y plane that are constant throughout the z dimension.

\n

I guess one way is to apply the same 2d transformation to every slice but would this be slow? Another approach could be to use scipy.ndimage.rotate but I'm worried about memory/speed since I think this requires numpy arrays as input.

", "messages": [ { "id": "1917012", "postDate": "08/28/2022 10:51:23", "content": "

Is anyone aware of a library that performs data augmentations on 3d volumes, compatible with PyTorch and ideally works on GPU?

\n

During training, I'm loading (B,C,D,H,W)=(batch_size,1,224,224,224) tensors and I'd like to apply rotations/flips in the x-y plane that are constant throughout the z dimension.

\n

I guess one way is to apply the same 2d transformation to every slice but would this be slow? Another approach could be to use scipy.ndimage.rotate but I'm worried about memory/speed since I think this requires numpy arrays as input.

", "rawMarkdown": "Is anyone aware of a library that performs data augmentations on 3d volumes, compatible with PyTorch and ideally works on GPU? \n\nDuring training, I'm loading (B,C,D,H,W)=(batch_size,1,224,224,224) tensors and I'd like to apply rotations/flips in the x-y plane that are constant throughout the z dimension. \n\nI guess one way is to apply the same 2d transformation to every slice but would this be slow? Another approach could be to use [scipy.ndimage.rotate](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.rotate.html) but I'm worried about memory/speed since I think this requires numpy arrays as input.", "votes": null }, { "id": "1917077", "postDate": "08/28/2022 12:06:22", "content": "

You could check out Kornia: https://kornia.readthedocs.io/en/latest/augmentation.module.html#transforms3d

", "rawMarkdown": "You could check out Kornia: https://kornia.readthedocs.io/en/latest/augmentation.module.html#transforms3d", "votes": null }, { "id": "1917107", "postDate": "08/28/2022 12:36:04", "content": "

Just what I was looking for, thank you!

", "rawMarkdown": "Just what I was looking for, thank you!", "votes": null }, { "id": "1917176", "postDate": "08/28/2022 13:37:36", "content": "

Here is my recommendation: https://github.com/ZFTurbo/volumentations

", "rawMarkdown": "Here is my recommendation: https://github.com/ZFTurbo/volumentations", "votes": null }, { "id": "1918506", "postDate": "08/29/2022 16:12:15", "content": "

I would suggest MONAI (https://monai.io) which is developed from medical DL projects, it has plenty of augmentation strategies, not limited to flip/rotate.

", "rawMarkdown": "I would suggest MONAI (https://monai.io) which is developed from medical DL projects, it has plenty of augmentation strategies, not limited to flip/rotate.", "votes": null }, { "id": "1922184", "postDate": "09/01/2022 10:22:23", "content": "

(B,C,D,H,W)=(batch_size,1,224,224,224)
\nWhat's your train batch size? And What's your GPU type and how many GPUs you have?

", "rawMarkdown": "(B,C,D,H,W)=(batch_size,1,224,224,224)\nWhat's your train batch size? And What's your GPU type and how many GPUs you have?", "votes": null } ], "comments": [ { "id": 1917077, "author_name": "surajghuwalewala", "author_url": "", "post_date": "08/28/2022 12:06:22", "content": "

You could check out Kornia: https://kornia.readthedocs.io/en/latest/augmentation.module.html#transforms3d

", "votes": null, "replies": [ { "id": 1917107, "author_name": "samuelcortinhas", "author_url": "", "post_date": "08/28/2022 12:36:04", "content": "

Just what I was looking for, thank you!

", "votes": null, "replies": [] } ] }, { "id": 1917176, "author_name": "analokamus", "author_url": "", "post_date": "08/28/2022 13:37:36", "content": "

Here is my recommendation: https://github.com/ZFTurbo/volumentations

", "votes": null, "replies": [] }, { "id": 1918506, "author_name": "bardiakh", "author_url": "", "post_date": "08/29/2022 16:12:15", "content": "

I would suggest MONAI (https://monai.io) which is developed from medical DL projects, it has plenty of augmentation strategies, not limited to flip/rotate.

", "votes": null, "replies": [] }, { "id": 1922184, "author_name": "h030162", "author_url": "", "post_date": "09/01/2022 10:22:23", "content": "

(B,C,D,H,W)=(batch_size,1,224,224,224)
\nWhat's your train batch size? And What's your GPU type and how many GPUs you have?

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1917012": "Is anyone aware of a library that performs data augmentations on 3d volumes, compatible with PyTorch and ideally works on GPU? \n\nDuring training, I'm loading (B,C,D,H,W)=(batch_size,1,224,224,224) tensors and I'd like to apply rotations/flips in the x-y plane that are constant throughout the z dimension. \n\nI guess one way is to apply the same 2d transformation to every slice but would this be slow? Another approach could be to use [scipy.ndimage.rotate](https://docs.scipy.org/doc/scipy/reference/generated/scipy.ndimage.rotate.html) but I'm worried about memory/speed since I think this requires numpy arrays as input.", "1917077": "You could check out Kornia: https://kornia.readthedocs.io/en/latest/augmentation.module.html#transforms3d", "1917107": "Just what I was looking for, thank you!", "1917176": "Here is my recommendation: https://github.com/ZFTurbo/volumentations", "1918506": "I would suggest MONAI (https://monai.io) which is developed from medical DL projects, it has plenty of augmentation strategies, not limited to flip/rotate.", "1922184": "(B,C,D,H,W)=(batch_size,1,224,224,224)\nWhat's your train batch size? And What's your GPU type and how many GPUs you have?" }, "source": "meta" }