tinydiff: A tiny diffusion training example

Reference:

• smalldiffusion: https://github.com/yuanchenyang/smalldiffusion
• https://sander.ai/2022/05/26/guidance.html

Classifier-free guidance

Train a conditional diffusion model p(x|y) with conditional dropout: some percentage of time(10-20% tends to work well), the conditioning information y is removed. In practice, it is often replaced with a special input value representing the absence of conditioning information. The resulting model is now able to function both as conditional model p(x|y) and as an unconditional model p(x) depending on whether the conditioning signal is provided.