Extract values from a 1-D numpy array for a batch of indices. :param arr: the 1-D numpy array. :param timesteps: a tensor of indices int

Evaluates how good the adversarial examples are ARGS: ground_truths: Variable (NxCxHxW) - examples before we did

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

Make a standard normalization layer. :param channels: number of input channels. :return: an nn.Module for normalization.

Logs the data point by specifying each of epoch, minibatch, value ARGS: name : string - name of existing series in self.series

Write text to stdout (and a file) and optionally flush.

Create a 1D, 2D, or 3D convolution module.

The same initialization used in DDPM.

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

Flush written text to both stdout and a file, if open.

Get activation functions from the config file.

Returns some grid parameters such that the minibatch of images isn't changed when forward is called on it ARGS: shape

Write the sequence of args, with no separators, to the console and output files (if you've configured an output file).

Take the mean over all non-batch dimensions.

FullSpatial just has parameters that are the grid themselves. Forward then will just call grid sample using these params directly

Factory class to generate per_minibatch instances of Adversarial perturbations. ARGS: perturbation_class : class - su

Little helper method to get the tensors of the adversarial images directly

Convert the torso of the model to float16.

3x3 convolution with DDPM initialization.

Create a linear module.

Defaults for image training.

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

Outputs string name of file. ARGS: experiment_name : string - name of experiment we're saving architecture : string - abbreviatio

Does some cleanup stuff after we finish on a minibatch: - clears the fixed images for ReferenceRegularizers - zeros grads - c

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

Obtain normalization modules from the config file.

Parameters to determine the marginal distribution of the SDE, $p_t(x)$.

Returns the 'norm' of this transformation in terms of an LP norm on the parameters, summed across each transformation per minibatch

Apply the model to get p(x_{t-1} | x_t), as well as a prediction of the initial x, x_0. :param model: the model, which takes

Compute the mean and variance of the diffusion posterior: q(x_{t-1} | x_t, x_0)

Zero out the parameters of a module and return it.

Construct an SDE. Args: N: number of discretization time steps.

Implements the function f(n) defined here: https://en.wikipedia.org/wiki/HSL_and_HSV#Alternative_HSV_to_RGB

Get a term for the variational lower-bound. The resulting units are bits (rather than nats, as one might expect). This allow

Flush, close possible files, and remove stdout/stderr mirroring.

Extract coefficients from a based on t and reshape to make it broadcastable with x_shape.

Converts an Nx3xWxH tensor in RGB color space to a Nx3xWxH tensor in this color space. All outputs should be in the 0-1 range.

Returns a flax optimizer object based on `config`.

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

Returns parameters for identity affine transformation ARGS: shape: torch.Size - shape of the minibatch of images we'll be

Importance-sample timesteps for a batch. :param batch_size: the number of timesteps. :param device: the torch device to save

Setup before calling loss on a new minibatch. Ensures the correct fix_im for reference regularizers and that all grads are zeroed

Converts an Nx3xWxH tensor in this color space to a Nx3xWxH tensor in RGB color space.