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.

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

r""" Return the parameters of models corresponding to keys If keys are empty, return the parameters of all models Args:

The same initialization used in DDPM.

Get activation functions from the config file.

3x3 convolution with DDPM initialization.

r""" Args: keys: the keys of objects If empty, return the states of all objects

r""" Args: states: a dict of states of objects keys: the keys of objects If empty, load states for all

Zero out the parameters of a module and return it.

Build pretrained InceptionV3 Parameters ---------- output_blocks : list of int Indices of blocks to return featur

Obtain normalization modules from the config file.

r""" Args: dataset: an instance of Dataset score_fn: a batch of data -> a batch of scalars batch_size: the batch size

r""" Args: shift1: whether to shift the index

r""" Manage a dict of objects

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

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

Create a linear module.

Create sinusoidal timestep embeddings. :param timesteps: a 1-D Tensor of N indices, one per batch element. These may b

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

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

A fast approximation of the cumulative distribution function of the standard normal.

Calculates the FID of two paths

Evaluate a function without caching intermediate activations, allowing for reduced memory at the expense of extra compute in the backward pas

r""" Exponential moving average theta <- beta * theta + (1 - beta) * theta_src Args: src: the source model

r""" The Euler Maruyama sampler for reverse SDE / ODE See `Score-Based Generative Modeling through Stochastic Differential Equations`