Augmentations in batch-size level

[black0017]

Hello @fepegar , really great work!

One short question: is it possible to apply the implemented augmentations on a batch-level (5D tensor)?

Quick example

import torch
import torchio.transforms as T
x = torch.rand(16, 1, 1, 256, 256)
transf = T.RandomElasticDeformation() # or # transf = T.Crop(cropping=(0, 10, 10)) 
y = transf(x)

Results in:

ValueError: The input must be a 4D tensor with dimensions (channels, x, y, z) but it has shape (16, 1, 1, 256, 256)

Why is it not possible to apply transformations on a batch-level? Is there any quick fix to it?

Thanks in advance,
Nikolas

[fepegar]

Hi, Nikolas. Thanks for your kind words!

What about just iterating over your batch?

import torch
import torchio.transforms as T
x = torch.rand(16, 1, 1, 256, 256)
transf = T.RandomElasticDeformation()
y = torch.stack([transf(instance) for instance in x])

[black0017]

Nice, I see. Thanks a lot for the reply @fepegar !!!

[fepegar]

No worries!