Multi Head Attention in Vision Transformers: Explanation and Full Implementation
This video covers everything about self attention in Vision Transformer - VIT , and its implementation from scratch.
I go over all the details and explain everything happening inside attention in vision transformer in detail through visualizations and also go over how an implementation of self-attention from scratch would look like in Pytorch.
I cover Vision transformer ( VIT ) in three parts:
1. Patch Embedding in Vision Transformer VIT - • PATCH EMBEDDING | Vision Transformers expl...
2. Self Attention in Vision Transformer VIT - This video
3. Building Vision Transformer and visualizations - • Image Classification Using Vision Transfor...
Paper Link - https://tinyurl.com/exai-vit-paper
Implementation - https://tinyurl.com/exai-vit-code
Other Good Resources
Yannic Kilcher | An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Paper Explained) - • An Image is Worth 16x16 Words: Transformer...
AI Coffee Break with Letitia | An image is worth 16x16 words: ViT | Vision Transformer explained - • An image is worth 16x16 words: ViT | Visio...
James Briggs | Vision Transformers (ViT) Explained + Fine-tuning in Python - • Vision Transformers (ViT) Explained + Fine...
Good Place to understand general transformer further - https://tinyurl.com/exai-vit-transformer
TimeStamps :
00:00 Intro
00:33 Intuition of What isAttention & Why its helpful
03:23 Inside Attention - What is Relevant
07:53 Inside Attention - Building Context Representation
08:45 Building Context Representation For All Patches
09:45 Why Multi Head Attention
11:15 Building Context Representation For Multi Head Attention
12:35 Combining Wq, Wk,Wv matrix
13:34 Shapes of Every Matrix in Attention
14:48 Implementation Parts of Attention
15:12 Pytorch Implementation for Attention in Vision Transformer VIT
18:26 Outro
Subscribe to Channel - https://tinyurl.com/exai-channel-link
Background Track - Fruits of Life by Jimena Contreras
Email - [[email protected]](mailto:[email protected])