🤖 Yet Another Transformer Implementation 🤖

Warning

This repository was developed for academic and personal purposes in order to better understand the underlying architecture of the Transformer and to use it for future small projects.

---

📦 Installation

You can install the package by either:

• using pip

  pip install git+https://github.com/RistoAle97/yati

  This will not install the dev dependencies listed in pyproject.toml.

• cloning the repository and installing the dependencies

  git clone https://github.com/RistoAle97/yati
  
  pip install -e yati
  pip install yati[dev]  # if you want to contribute to this project

---

🛠️ Implementation details

Note

Some implementation choices have been made that may differ from the original paper [1]:

• The source and target embeddings are shared, so a unified vocabulary (one for all the languages in a NMT task to give an example) is needed.
• The embeddings are tied to the linear output (i.e.: they share the weights).
• Pre-normalization was employed instead of post-normalization [2].
• Layer normalization [3] is performed at the end of both the encoder and decoder stacks.
• There is no softmax layer as it is already used by the CrossEntropy loss function implemented in PyTorch.

Hereafter a comparison between the original transformer and the one from this repository.

Original	This repository

---

📚 Bibliography

[1] Vaswani, Ashish, et al. "Attention is all you need.". Advances in neural information processing systems 30 (2017).

[2] Nguyen, Toan Q., and Julian Salazar. "Transformers without tears: Improving the normalization of self-attention.". arXiv preprint arXiv:1910.05895 (2019).

[3] Ba, Jimmy Lei, Jamie Ryan Kiros, and Geoffrey E. Hinton. "Layer normalization". arXiv preprint arXiv:1607.06450 (2016).

Some additional nice reads:

• Transformer Details Not Described in The Paper.

📝 License

This project is MIT licensed.