Hi, congratulations on your excellent job!
|
def upsampling_fn(x, alpha, sigma, dataset): |
|
base_resolution = base_resolutions[dataset] |
|
curr_resolution = x.shape[-1] |
|
n = int(np.log2(base_resolution / curr_resolution)) |
|
x = repeat(upsample, x, n) |
|
std = np.sqrt(sigma**2 + alpha**2 * stds[dataset][curr_resolution]**2) |
|
noise = torch.normal(mean=0, std=std, size=x.shape, device=x.device) |
|
noise = noise - repeat(upsample, repeat(downsample, noise, n), n) |
|
return x + noise |
I find that the upsampling_fn just increases the resolution of the input. I want to know whether the PCA analysis is useful in this process.
