q_sample results look different to paper

[mattia-git]

Hi,

I am trying to use your code to apply the noising forward process to a binary segmentation mask. Following the same setup as the paper (schedule=linear, T=1000, background=0, ROI=1, shape=(N, 1, H, W)), I get the following noising process which looks different to the output shown in your paper:

Is this the expected output, or am I doing something wrong?

[takimailto]

Hello, thank you for your attention, can you give the code？

[LiangBinCams]

I had the same problem. Did you fix it?

[LiangBinCams]

Hi,

I am trying to use your code to apply the noising forward process to a binary segmentation mask. Following the same setup as the paper (schedule=linear, T=1000, background=0, ROI=1, shape=(N, 1, H, W)), I get the following noising process which looks different to the output shown in your paper:

Is this the expected output, or am I doing something wrong?

I had the same problem. Did you fix it?

[LiangBinCams]

Hi, I am trying to use your code to apply the noising forward process to a binary segmentation mask. Following the same setup as the paper (schedule=linear, T=1000, background=0, ROI=1, shape=(N, 1, H, W)), I get the following noising process which looks different to the output shown in your paper: Is this the expected output, or am I doing something wrong?

I had the same problem. Did you fix it?

I figured out the problem. The reason is here:

def q_mean(self, x_start, t):
    """
    Get the distribution q(x_t | x_0).

    :param x_start: the [N x C x ...] tensor of noiseless inputs.
    :param t: the number of diffusion steps (minus 1). Here, 0 means one step.
    :return: Binomial distribution parameters, of x_start's shape.
    """
    mean = _extract_into_tensor(self.alphas_cumprod, t, x_start.shape) * x_start 
    + (1 - _extract_into_tensor(self.alphas_cumprod, t, x_start.shape)) / 2

    return mean

this function returns is only _extract_into_tensor(self.alphas_cumprod, t, x_start.shape) * x_start !!