Tweak quality reduction formula

Suggestion:
TorchIO-project/torchio#765 (comment)

miqa/learning/nn_training.py

@@ -227,11 +227,11 @@ def apply_transform(self, subject: torchio.Subject) -> torchio.Subject:
 
             # now determine how much quality was reduced
             applied_params = transformed_subject.applied_transforms[-1][1]
-            times = applied_params['times']['img']
+            time = applied_params['times']['img']
             degrees = np.sum(np.absolute(applied_params['degrees']['img']))
             translation = np.sum(np.absolute(applied_params['translation']['img']))
             # motion earlier in the acquisition process produces more noticeable artifacts
-            quality_reduction = clamp(degrees + translation, 0, 10) * (1.0 - times)
+            quality_reduction = clamp(degrees + translation, 0, 10) * min(time, 1.0 - time)
 
             # update the ground truth information
             new_quality = original_quality - quality_reduction
@@ -305,8 +305,7 @@ def create_train_and_test_data_loaders(df, count_train):
     rescale = torchio.RescaleIntensity(out_min_max=(0, 1))
     ghosting = CustomGhosting(p=0.3, intensity=(0.2, 0.8))
     motion = CustomMotion(p=0.2, degrees=5.0, translation=5.0, num_transforms=1)
-    # transforms = torchio.Compose([rescale, ghosting, motion])
-    transforms = torchio.Compose([rescale, motion])
+    transforms = torchio.Compose([rescale, ghosting, motion])
 
     # create a training data loader
     train_ds = torchio.SubjectsDataset(train_files, transform=transforms)