Fix Blur using only the first standard deviation (#772)

* Fix Blur using only the first standard deviation Fixes #769. * Use the same Gaussian kernel for all channels
TorchIO-project · Dec 7, 2021 · 68755fb · 68755fb
1 parent 3cebdec
commit 68755fb
Showing 1 changed file with 8 additions and 9 deletions.
diff --git a/torchio/transforms/augmentation/intensity/random_blur.py b/torchio/transforms/augmentation/intensity/random_blur.py
@@ -5,7 +5,6 @@
 import numpy as np
 import scipy.ndimage as ndi
 
-from ....utils import to_tuple
 from ....typing import TypeData, TypeTripletFloat, TypeSextetFloat
 from ....data.subject import Subject
 from ... import IntensityTransform
@@ -40,9 +39,9 @@ def __init__(
 
     def apply_transform(self, subject: Subject) -> Subject:
         arguments = defaultdict(dict)
-        for name, image in self.get_images_dict(subject).items():
-            stds = [self.get_params(self.std_ranges) for _ in image.data]
-            arguments['std'][name] = stds
+        for name in self.get_images_dict(subject):
+            std = self.get_params(self.std_ranges)
+            arguments['std'][name] = std
         transform = Blur(**self.add_include_exclude(arguments))
         transformed = transform(subject)
         return transformed
@@ -72,15 +71,15 @@ def __init__(
         self.args_names = ('std',)
 
     def apply_transform(self, subject: Subject) -> Subject:
-        std = self.std
+        stds = self.std
         for name, image in self.get_images_dict(subject).items():
             if self.arguments_are_dict():
-                std = self.std[name]
-            stds = to_tuple(std, length=len(image.data))
+                stds = self.std[name]
+            stds_channels = np.tile(stds, (image.num_channels, 1))
             transformed_tensors = []
-            for std, tensor in zip(stds, image.data):
+            for std, channel in zip(stds_channels, image.data):
                 transformed_tensor = blur(
-                    tensor,
+                    channel,
                     image.spacing,
                     std,
                 )