Blur image directly after inference on GPU

vision
1

Is there any way to run blurring on a torch tensor on a GPU? I am running a segmentation model and would like to blur the segmented areas of the original image directly on the gpu afterwards (not put it onto the CPU first to blur).

2

Does torchvision.transforms.GaussianBlur do what you need?

import torch, torchvision
device = torch.device("cuda")
tensor = torch.rand(3, 224, 224).to(device)
blurred_tensor = torchvision.transforms.GaussianBlur(3)(tensor)
print(blurred_tensor.device)

Output:
cuda:0
3

Cool thanks you! Is the blurring operation also done on the gpu?

4

I believe so. Perhaps @ptrblck can you confirm?

1 Like
5

Yes, you are right and internally this method will be used, which will apply a conv2d on the input tensor using the specified img.device arguments.
If a PIL.Image is passed it’ll be transformed to a tensor before the blur transformation is applied and the result will then be transformed back to a PIL.Image as seen here (so the CPU will be used in this case).

1 Like
6

Cool thank you very much!