Convert tensor into -1 to 1 range

vision
1

I want to do image restoration, restore_image = (orig_image - B)/C.
orig_image is in the range (-1,1) but B is in the range of (0,1) and C is random output of LeakyRelU (-infinity, + infinity).

I want to convert B and C in the range (-1,1) for image restoration.

2

tensor = (tensor/torch.max(tensor) -0.5 ) * 2, should get the job done

3

Is this formula remain same if the negative value is higher than positive means minimum value is -30 and maximum value is 10?

4

I gave that solution assuming you were doing it for regular images. Anyways, for C, the following might be a better option.

tensor = (tensor/torch.max(torch.abs(tensor))

This itself should squish the values between - 1 and +1. For B (whose values are between 0 to 1), you can use the previous formula

5

I restored the image using, restore_image = (orig_image - B)/C.

Here C is output of training. The value comes in the range of (-2,10). After that i apply the transformation to get the value between (-1,1) as mentioned. After transformation,C values are in the range tensor([[[[-1.0349e-02, 4.3450e-02, 3.3275e-02, …, 3.0451e-02, etc.

So when i divide (orig_image - B) by C, the output values are tensor([[[[ 4.2118e+02, -1.7778e+01, -1.5082e+02, …, -2.7659e+02, -4.8812e+01, -1.6656e+01] etc.

Again I transformed it in the range of (-1,1) and I got the final value tensor([[[[ 2.9290e-04, -1.2363e-05, -1.0488e-04, …, -1.9235e-04, -3.3945e-05, -1.1583e-05] etc.

This output restored image is constant with some gray level.
Divide by fraction is creating issue. Can you throw some light to solve the issue?

6

You are displaying the image with the pixel values lying lie within (-1,1). Try converting the output back between (0-255) before displaying it.
For values between -1 and 1

img = (tensor/2 + 0.5) * 255
img = img.detach().numpy()
img = img.astype(np.uint8)

Then display the image