This post is divided into six parts; they are: • Why Transformer is Better than Seq2Seq • Data Preparation and Tokenization • Design of a Transformer Model • Building the Transformer Model • Causal Mask and Padding Mask • Training and Evaluation Traditional seq2seq models with recurrent neural networks have two main limitations: • Sequential processing prevents parallelization • Limited ability to capture long-term dependencies since hidden states are overwritten whenever an element is processed The Transformer architecture, introduced in the 2017 paper “Attention is All You Need”, overcomes these limitations.
