Hi! I’m doing image Super-resolution with Celeba dataset, but I always got theses images with noise pixels which are very unnatural in face images.
No matter how many epoch trained, I still can not get rid of these noise pixels !
I use Pillow to read images then convert them to RGB form, and normalize them in this way:
self.transforms_HR = transforms.Compose([
transforms.Normalize((0.5, 0.5, 0.5),
(0.5, 0.5, 0.5))])
As I use the Visdom to navigate the training process, I transform the tensor in this way:
image_numpy = (np.transpose(image_numpy, (1, 2, 0)) + 1) / 2.0 * 255.0
BTW the image_numpy is a image grid with shape of (3, H, W) output by a function in torchvision
So I wonder if there is a way to get rid of these noise pixel?