Please note the Resize, Crop usage of torchvision.transforms!

[kmkmkr]

Differences between Resize and Crop arguments may cause incorrect comparisons between corruption data and clean data.

The label of this image is tench, in the case of this image, method 1 clearly improves the accuracy of the original (severity=0)
In the transformation of the original image, method1 is incorrect

For a correct comparison, I would suggest using method2 .

clean_image_path : val/n01440764/ILSVRC2012_val_00041939.JPEG.png
method1

transforms.Resize((256, 256)),
transforms.CenterCrop((224, 224)),

method2

transforms.Resize(256),
transforms.CenterCrop(224),

method1 result (severity:0->1->2->3->4->5)

method2 result(severity:0->1->2->3->4->5)