filter_visualizer (and img reference)

[RR5555]

Greetings,
I adapted your code in ''pure'' pytorch and replaced the calls to cv2 by calls to PIL.
In doing so, I bumped into what could cause an issue in the original code (though I did never run the original code, so maybe it does not), in the hook, I believe that:
self.features = torch.tensor(output,requires_grad=True).cuda()
could actually break the backward of the loss.
Replacing it by:
self.features = output
did solve the problem.

A detail that could be nice to change would be to add the picture used in the Calculate... notebook in the "pictures" directory and modify the corresponding call:
picture = PIL.Image.open("test/206px-Wattledcranethumb.jpg")

PS: Thank you very much for your post on towardsdatascience and the associated code.

[leopardyao]

beautiful code but the above issue is also exactly the problem which I am running into
pytorch/pytorch#42188