4
$\begingroup$

I am trying to implement model described in Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network in which author says in section 3.2 that

We scaled the range of the LR input images to [0, 1] and for the HR images to [−1, 1]. The MSE loss was thus calculated on images of intensity range [−1, 1]

Now I know that when Gray-scaled images are represented in float it's value is in range [0,1] but how do you scale your colour images in range of [0,1] and [-1,1]? And why authors are using two different scales for input and High resolution images?

Any insights would be helpful.

$\endgroup$
1
$\begingroup$

You can use the following scalings

$$x’=\dfrac{x}{255} \qquad (1)$$ $$x’=\dfrac{x-127.5}{127.5} = \dfrac{x}{127.5}-1 \qquad (2)$$

for rescaling to $[0,1]$ or $[-1,1]$.

The rescaling of inputs tries to keep the range of weights in a small range. Theoretically it is not necessary to rescale your inputs because it can be compensated by an appropriate redefinition of your weights. Practically it is important because your weights might occupy a very large range of values.

In order to understand this I will construct a toy example with two inputs $x_1= 1$ and $x_2=1$ and a simple linear regression $y_n=w_1x_{1n}+w_2x_{2n}$. Assume the true values for the weight are $w_1=1$ and $w_2=1$. Now, assume we have the same data set but with different scales for $x_1$ and $x_2$ such that $x_1=10^k$ and $x_2=10^{-k}$. In order to obtain the same inputs we would need to have inverse weights resulting in $w_1=10^{-k}$ and $w_2=10^k$. Theoretically this is not a big deal but practically we see that for large values of $k$ we will need to use variables that store numerical values between $10^{-k}$ and $10^k$.

I don’t know the reason why the authors of the cited paper used specific transformations. Both should lead to the almost same performance as the bias should be able compensate for the additional $-1$ and the reacaling could be compensated by the weights for each input. There might be differences because of the stochasticity of many optimization techniques and because the final solution is very likely not the global optimum. I would choose one over the other if I want to compare my model with a specific model from a paper. In such a case I would use the same transformation as applied in the paper.

You could also use different scales in order to better discriminate the type of input (low resolution vs high resolution) by looking at the transformed values.

| improve this answer | |
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.