Combined Affine/Elastic augmentations #1052
Labels
enhancement
New feature or request
Comments
|
Hi, @cfhammill. It would be a nice feature. Would you be available to contribute such a transform? |
mrdkucher
added a commit
to mrdkucher/torchio
that referenced
this issue
Apr 19, 2023
SimpleITK's ResampleImageFilter is an expensive operation, especially when its called sequentially for both affine and elastic transformations. By combining the SimpleITK transforms for both augmentations, the processing time for a sample can be reduced. Currently the augmentation uses a combined probability for applying the transform, rather than independently applying them. Resolves: TorchIO-project#1052
8 tasks
mrdkucher
added a commit
to mrdkucher/torchio
that referenced
this issue
Apr 20, 2023
SimpleITK's ResampleImageFilter is an expensive operation, especially when its called sequentially for both affine and elastic transformations. By combining the SimpleITK transforms for both augmentations, the processing time for a sample can be reduced. Currently the augmentation uses a combined probability for applying the transform, rather than independently applying them. Resolves: TorchIO-project#1052
mrdkucher
added a commit
to mrdkucher/torchio
that referenced
this issue
Aug 25, 2023
SimpleITK's ResampleImageFilter is an expensive operation, especially when its called sequentially for both affine and elastic transformations. By combining the SimpleITK transforms for both augmentations, the processing time for a sample can be reduced. Currently the augmentation uses a combined probability for applying the transform, rather than independently applying them. Resolves: TorchIO-project#1052
mrdkucher
pushed a commit
to mrdkucher/torchio
that referenced
this issue
Aug 27, 2023
SimpleITK's ResampleImageFilter is an expensive operation, especially when its called sequentially for both affine and elastic transformations. By combining the SimpleITK transforms for both augmentations, the processing time for a sample can be reduced. Currently the augmentation uses a combined probability for applying the transform, rather than independently applying them. Add combined affine and elastic deformation augmentation SimpleITK's ResampleImageFilter is an expensive operation, especially when its called sequentially for both affine and elastic transformations. By combining the SimpleITK transforms for both augmentations, the processing time for a sample can be reduced. Currently the augmentation uses a combined probability for applying the transform, rather than independently applying them. Resolves: TorchIO-project#1052 Remove duplicated code Update pytests
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
🚀 Feature
A combined elastic/affine augmentation.
Motivation
Currently CPU use is training bottleneck for training a relatively large vision transformer, most of the time is spent in augmentation on sitk's resampling code, it looks like resampling is getting triggered once for each of the random affine and random elastic transformations.
Pitch
Two additional operations RandomAffineThenElastic, and RandomElasticThenAffine (I would use the later), where the two random transformations are merged and use only a single resampling.
Notes
Thanks for the great package!
The text was updated successfully, but these errors were encountered: