Transformer Encoder in 100 Lines of Code

Dive into a comprehensive 50-minute video tutorial that breaks down the Transformer Encoder architecture into 100 lines of code. Learn about word embeddings, attention heads, dropout, data batching, and the intricacies of the encoder layers. Explore key concepts such as multi-head attention, layer normalization, and feed-forward networks. Gain hands-on experience with PyTorch implementations, including nn.Module and nn.Sequential. Understand the flow of data through the encoder and discover why certain components are crucial for the transformer's performance. Perfect for those looking to deepen their understanding of natural language processing and deep learning architectures.

What we will cover
Introducing Colab
Word Embeddings and d_model
What are Attention heads?
What is Dropout?
Why batch data?
How to sentences into the transformer?
Why feed forward layers in transformer?
Why Repeating Encoder layers?
The “Encoder” Class, nn.Module, nn.Sequential
The “EncoderLayer” Class
What is Attention: Query, Key, Value vectors
What is Attention: Matrix Transpose in PyTorch
What is Attention: Scaling
What is Attention: Masking
What is Attention: Softmax
What is Attention: Value Tensors
CRUX OF VIDEO: “MultiHeadAttention” Class
Returning the flow back to “EncoderLayer” Class
Layer Normalization
Returning the flow back to “EncoderLayer” Class
Feed Forward Layers
Why Activation Functions?
Finish the Flow of Encoder
Conclusion & Decoder for next video