How to generate images like Class Specific Image Generation section about pytorch-cnn-visualizations HOT 3 CLOSED

Hello,

That example uses an image generated from a random uniform distribution between 0 and 255.
This line:
self.created_image = np.uint8(np.random.uniform(0, 255, (224, 224, 3)))

Running the code directly produces following images for me (AlexNet - flamingo):
Iter 0:

Iter 100:

Iter 200:

I used VGG19 to generate the snake and the spider. I think I also incorporated a Gaussian blur for that (unrelated to the generation technique, just a way to improve the quality of image).

Perhaps I should include the code for VGG19, incorporating with blurring as well. Though, technically, it is just passing VGG19 as the model instead of AlexNet and blurring the generated image every n iteration.

Would it be helpful ?

Edit: Also don't forget the random nature of the process, two generated images are very unlikely to be the same. (Although they should be similar to a degree)

from pytorch-cnn-visualizations.

Thanks, I get the similar result using gaussian blur.

from pytorch-cnn-visualizations.

No problem, I will update the documentation as well so that its clear for everyone.

from pytorch-cnn-visualizations.