3x3 convolution with PyTorch initialization. Same as NCSNv1/NCSNv2.

Compute the output of the score-based model. Args: x: A mini-batch of input data. labels: A mini-batch of conditioning va

r""" Equivalent to `torch.randn_like` but supports a pseudorandom number generator argument. :param int seed: the seed for the random

Copy current parameters into given collection of parameters. Args: parameters: Iterable of `torch.nn.Parameter`; the parameters to be

The same initialization used in DDPM.

Get activation functions from the config file.

Update currently maintained parameters. Call this every time the parameters are updated, such as the result of the `optimizer.step()` ca

Calculate Dg(x) the gradient of the regularizer g at input x :param x: torch.tensor Input image :param sigma: Denoiser level

3x3 convolution with DDPM initialization.

Flatten a torch tensor `x` and convert it to numpy.

The drift function of the reverse-time SDE.

Obtain normalization modules from the config file.

Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters without affecting the original op

Save the current parameters for restoring later. Args: parameters: Iterable of `torch.nn.Parameter`; the parameters to be temp

Denoising with Gradient Step Denoiser :param x: torch.tensor input image :param sigma: Denoiser level (std) :return:

Get sigmas --- the set of noise levels for SMLD from config files. Args: config: A ConfigDict object parsed from the config file Returns

r""" Sets the seed for the random number generator. :param int seed: the seed to set for the random number generator. If not provided

Wraps `torchvision.models.inception_v3` Skips default weight inititialization if supported by torchvision version. See https://github.com/mse

Calculation of the statistics used by the FID. Params: -- images : Images -- model : Instance of inception model -- batch_s

s-fold downsampler Keeping the upper-left pixel for each distinct sfxsf patch and discarding the others x: tensor image, NxCxWxH

Form a torch tensor with the given `shape` from a flattened numpy array `x`.

Generate a 2D Gaussian kernel.

Create the divergence function of `fn` using the Hutchinson-Skilling trace estimator.

Compute an unbiased estimate to the log-likelihood in bits/dim. Args: model: A score model. data: A PyTorch tensor. Returns:

Compute the loss function. Args: model: A velocity model. batch: A mini-batch of training data. Returns:

Black mask that looks like paintbrush on the image. Make it random

Random mask on x

Black square mask of 20 x 20 pixels at the center of the image

r""" Updates the standard deviation of the noise. :param float, torch.Tensor sigma: standard deviation of the noise.

s-fold upsampler Upsampling the spatial size by filling the new entries with zeros x: tensor image, NxCxWxH

copied and modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/init.py#L337

Apply givens rotation to H columns :param H: :param cs: :param ss: :param j: :return:

Arnoldi iteration to find the j th l2-orthonormal vector compute the j-1 th column of Hessenberg matrix

r""" Bicubic filter. It has size (4*factor, 4*factor) and is defined as .. math:: \begin{equation*} w(x, y) = \begi

Args: a: element h in position j b: element h in position j+1 Returns: cosine = a / \sqrt{a^2 + b^2} sine = -

Numpy implementation of the Frechet Distance. The Frechet distance between two multivariate Gaussians X_1 ~ N(mu_1, C_1) and X_2 ~ N(mu_2, C_2

Returns a function that can be as a classifier function. Copied from tfgan but avoid loading the model each time calling _classifier_fn Args