Channel-wise self-attention block.

A sequential module that passes timestep embeddings to the children that support it as an extra input.

A residual block that can optionally change the number of channels. :param channels: the number of input channels. :param emb_channels:

InceptionC block patched for FID computation

An attention block that allows spatial positions to attend to each other. Originally ported from here, but adapted to the N-d case. http

InceptionA block patched for FID computation

r""" dx = [f(x, t) - g(t)^2 s(x, t)] dt

r""" The forward process is q(x_0, x_1, ..., x_N), which is indexed from 0 to N Some codes use different indexes, such as q(x_-1, x_0,

Pretrained InceptionV3 network returning feature maps

`LSUN <https://www.yf.io/p/lsun>`_ dataset. Args: root (string): Root directory for the database files. classes (string or l

r""" E[xs|xt] = E[ E[xs|xt,x0] |xt] = E[xs|xt,x0=E[x0|xt]] in DDPM or DDIM forward process

A downsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determin

First InceptionE block patched for FID computation

Second InceptionE block patched for FID computation

NCSN++ model

NCSN++ model

A module which performs QKV attention.

r""" dx = [f(x, t) - g(t)^2 s(x, t)] dt + g(t) dw

An upsampling layer with an optional convolution. :param channels: channels in the inputs and outputs. :param use_conv: a bool determini

Channel-wise self-attention block. Modified from DDPM.

r""" CIFAR10 dataset Information of the raw dataset: train: 40,000 val: 10,000 test: 10,000 shape: 3 * 32

r""" The forward process is q(x_[0,T]) n -> t = n * T / N especially, n=0 -> t=0, data n=N -> t=T

r""" The forward process is q(x_[0,T])

r""" train: 162,770 val: 19,867 test: 19,962 shape: 3 * width * width

Combine information from skip connections.

Conv2d layer with optimal upsampling and downsampling (StyleGAN2).

r""" Output dataset after two transformations to the raw data: 1. distribution transform (e.g. binarized, adding noise), often irreversible, a par

Linear layer with `default_init`.

Maintains (exponential) moving average of a set of parameters.

Gaussian Fourier embeddings for noise levels.

r""" Imagenet64 dataset Information of the raw dataset: train: 1,281,149 test: 49,999 shape: 3 * 64 * 64

r""" Mnist dataset Information of the raw dataset: train: 50,000 val: 10,000 test: 10,000 shape: 1 * 28 *