CODEWITHSUNDEEP
pythonpytorchtensor
Raguel
Raguel

Reputation: 625

How to normalize PIL image between -1 and 1 in pytorch for transforms.Compose?

So I have been trying to find a way to normalize some PIL image pixel values between -1 and 1. I searched through documentation and didn't find solution. Only normalization in documentation is transforms.Normalize which normalizes with mean and std. So I am stuck on how to do it. This is my code:

train_transform = transforms.Compose([
        transforms.RandomHorizontalFlip(p=0.5),
        transforms.Resize(40),
        transforms.RandomCrop(32),
        # Normalize(-1, 1)  # Something like that
])

How can I normalize tensor between 2 numbers? For example:

[[1,2,3]
[3,2,1]]

Between -1 and 1

[[-1, 0, 1],
[1, 0, -1]]

Upvotes: 2

Views: 8057

Answers (1)

Ioannis Nasios
Ioannis Nasios

Reputation: 8527

Try:

transform = transforms.Compose([
    transforms.RandomHorizontalFlip(p=0.5)
    transforms.Resize(40),
    transforms.RandomCrop(32),
    transforms.ToTensor(),
    transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))])

  • ToTensor will scale image pixels from [0,255] to [0,1] values

  • Normalize will first subtract the mean so [0,1] values will transform to range [-0.5,0.5], and then will divide with std and [-0.5,0.5] will reach [-1,1] range

Upvotes: 10

Related Questions