In this video I implement the Vision Transformer from scratch. It is very much a clone of the implementation provided in [ Ссылка ]. I focus solely on the architecture and inference and do not talk about training. I discuss all the relevant concepts that the Vision Transformer is using e.g. patch embedding, attention mechanism, layer normalization and many others.

My implementation: [ Ссылка ]
timm implementation: [ Ссылка ]
lucidrains implementation: [ Ссылка ]

00:00​​​ Intro
01:20​​​ Architecture overview
02:53​​ Patch embedding module
06:39 Attention module
07:22 Dropout overview
08:11​​​ Attention continued 1
10:50 Linear overview
12:10 Attention continued 2
14:35 Multilayer perceptron
16:07 Block module
17:02 LayerNorm overview
19:31 Block continued
20:44 Vision transformer
24:52 Verification
28:01 Cat forward pass
29:10 Outro

If you have any video suggestions or you just wanna chat feel free to join the discord server: [ Ссылка ]

Twitter: [ Ссылка ]

Credits logo animation
Title: Conjungation · Author: Uncle Milk · Source: [ Ссылка ] · License: [ Ссылка ] · Download (9MB): [ Ссылка ]

свернуть