Cuda code build error

[chengyu89527]

Hi how should I understand the yellow part at method B.What's the different between this trianle and ware X thanks

[onat-dalmaz]

It means that when sampling from the denoising distribution pθ(xt−k|xt, y), a trivial (but valid) sample is to take the output of the generator Gθ(xt, y, t) as a deterministic sample. In other words, xˆt−k = Gθ(xt, y, t), where the hat symbol denotes that this is a sample from the denoising distribution. This is because the denoising distribution is parameterized by the generator network Gθ, which takes as input xt, y, and t. The delta distribution is just a special case of the denoising distribution where the output is a single point with probability 1, and can be represented by the delta function δ. Therefore, taking xˆt−k as the generator output is equivalent to sampling from the delta distribution, which is what is meant by "δ(xt−k − Gθ(xt, y, t))". I hope this clarifies things for you.

[chengyu89527]

thanks for replaying, however I can't understand "the output is a single point with probability 1" >_<. please help me

[onat-dalmaz]

Instead of sampling from a probability distribution, to obtain x_hat, you just go ahead and take the generator's output as a deterministic sample. It is NOT a probability at all.