Lesson 19: Deep Learning Foundations to Stable Diffusion

(All lesson resources are available at course.fast.ai.) In this lesson, Jeremy introduces Dropout, a technique for improving model performance, and with special guests Tanishq and Johno he discusses Denoising Diffusion Probabilistic Models (DDPM), the underlying foundational approach for diffusion models. The lesson covers the forward and reverse processes involved in DDPM, as well as the implementation of a noise predicting model using a neural network. The team also demonstrate an alternative approach to the implementation and discuss ways to improve training speed.
0:00:00 - Introduction and quick update from last lesson
0:02:08 - Dropout
0:12:07 - DDPM from scratch - Paper and math
0:40:17 - DDPM - The code
0:41:16 - U-Net Neural Network
0:43:41 - Training process
0:56:07 - Inheriting from miniai TrainCB
1:00:22 - Using the trained model: denoising with “sample” method
1:09:09 - Inference: generating some images
1:14:56 - Notebook 17: Jeremy’s exploration of Tanishq’s notebook
1:24:09 - Make it faster: Initialization
1:27:41 - Make it faster: Mixed Precision
1:29:40 - Change of plans: Mixed Precision goes to Lesson 20
Many thanks to Francisco Mussari for timestamps and transcription.

[@faqeerhasnain]

Love Jeremy so much,Thank You

[@timandersen8030]

Why is the coding implementation for sampling of x_t-1 is different @1:06:25 from the algorithm 2 for sampling mentioned in the paper @53:50 ??

[@deepschoolai]

Does the course dive deeper into the architecture of the Unet model. I feel there's a lot of intricacies we are missing out on there.

[@deepschoolai]

46:03 I might be wrong here, but I don't think sigma is sqrt of beta. It is the square root of beta tilde which is NOT the value that has linearly spaced values. Getting this from section 3.2 of the DDPM paper.

[@laugh_n_share_life]

you are correct, this lecture was below their normal standard

[@deepschoolai]

@@laugh_n_share_life whoa, that is certainly not what I'm saying here. Just a simple question

[@rjScubaSki]

Greek letters for identifiers are a pointless distraction imo - the alternative isn't spelled out greek letter names, but meaningful names. "For coders" ...

[@scottnewcomer1835]

I love the way Tanishq described the forward pass. Good on ya! kzbin.info/www/bejne/f6XcgGupaZ2tmsU

[@SandeepSinghPlus]

I cannot find DDPM notebook in diffusion-nbs repo? Can somebody past the link for the same?