Book Image

Mastering Transformers

By : Savaş Yıldırım, Meysam Asgari- Chenaghlu
Book Image

Mastering Transformers

By: Savaş Yıldırım, Meysam Asgari- Chenaghlu

Overview of this book

Transformer-based language models have dominated natural language processing (NLP) studies and have now become a new paradigm. With this book, you'll learn how to build various transformer-based NLP applications using the Python Transformers library. The book gives you an introduction to Transformers by showing you how to write your first hello-world program. You'll then learn how a tokenizer works and how to train your own tokenizer. As you advance, you'll explore the architecture of autoencoding models, such as BERT, and autoregressive models, such as GPT. You'll see how to train and fine-tune models for a variety of natural language understanding (NLU) and natural language generation (NLG) problems, including text classification, token classification, and text representation. This book also helps you to learn efficient models for challenging problems, such as long-context NLP tasks with limited computational capacity. You'll also work with multilingual and cross-lingual problems, optimize models by monitoring their performance, and discover how to deconstruct these models for interpretability and explainability. Finally, you'll be able to deploy your transformer models in a production environment. By the end of this NLP book, you'll have learned how to use Transformers to solve advanced NLP problems using advanced models.
Table of Contents (16 chapters)
1
Section 1: Introduction – Recent Developments in the Field, Installations, and Hello World Applications
4
Section 2: Transformer Models – From Autoencoding to Autoregressive Models
10
Section 3: Advanced Topics

Working with efficient self-attention

Efficient approaches restrict the attention mechanism to get an effective transformer model because the computational and memory complexity of a transformer is mostly due to the self-attention mechanism. The attention mechanism scales quadratically with respect to the input sequence length. For short input, quadratic complexity may not be an issue. However, to process longer documents, we need to improve the attention mechanism that scales linearly with sequence length.

We can roughly group the efficient attention solutions into three types:

  • Sparse attention with fixed patterns
  • Learnable sparse patterns
  • Low-rank factorization/kernel function

Let's begin with sparse attention based on a fixed pattern next.

Sparse attention with fixed patterns

Recall that the attention mechanism is made up of a query, key, and values as roughly formulated here:

Here, the Score function, which is mostly softmax, performs...