Lesson 15: Deep Learning Foundations to Stable Diffusion

[@markozege]

This is amazing. Using excel and formula tracing to demonstrate what receptive field is is just brilliant. On top of great explanations of all concepts, each lecture is packed with useful practical tips and tricks. Big thanks to Jeremy and the team for making the world a better place!

[@MoetasimRady]

Great work as usual, just wanted to point out something that confused me. At 47:37 Jeremy mistakenly said that the second kernel applied in the first layer is a "left edge detector" when actually it is a right edge detector.

[@myfolder4561]

'Unfold' convolution operation (ie kernel sliding through input matrix) is actually straightforward, for those who don't want to read through the docs or detailed explanation see if below example helps: rethink kernel slide through input matrix as matrix calculation. eg (3,3) kernel, (28,28) input matrix (with each kernel to take sumprod and store the result in its center location) results in output (26,26) (The edges of the original input matrix reduced, given a stride=1 and no padding regular kernel sliding. A second similar convolution would reduce (26,26) to (24,24). Third one will reduce to (22,22) etc. You get the idea) Going back to the first convolution operation beginning of the example, (3,3) kernel slide through a (3,3) input-subset then sumproduct, is equivalent to -> input-subset.reshape(1,9) @ kernel.reshape(1,9).transpose() -> combining every input-subset or 'unfold' gives (26x26, 3x3) or (676, 9) input matrix (because there're 26x26 times of kernel slides, each represents a sumprod operation) -> input matrix @ kernel becomes (676,9) @ (9,1)

[@recontemplator]

Methodologically great explanation. Solution explained based on the known answer! Thank you.

[@michaelmuller136]

Great as always, i especially like the python tips, i always thought pip install -e stands for everything, editable makes much more sense!

[@ikaankeskin7473]

Where can we find the notebook?

[@SuperOnlyP]

I have the same question.

[@jorgbonfert4546]

💪 Promo'SM