Build Diffusion Models from Scratch

• The corruption process (blending noise to data)
• what a unet is , and how to implement an extremely minimal one from scratch
• diffusion model training
• sampling theory

Then compare our version with the diffusers DDPM implementation , exploring:

• Improvements over our mini unet
• DDPM noise schedule
• differences in training objective
• timestep conditioning
• sampling approaches

Syntax	Desddcriptiond
Header	Title
Paragraph	Text

Data preperation

using MNIST

dataset = torchvision.datasets.MNIST(
    root = "mnist/", train=True, download = True, transform = torchvision.transforms.ToTensor()
)


train_dataloader = DataLoader(dataset, batch_size=8, shuffle=True)

transform = torchvision.transforms.ToTensor convert a PIL image or ndarray to tensor and scale the value accordingly ToTensor doc

Corruption (adding noise)

use torch.rand_like