Transformer Backward Propagation

Introduction

This is an implementation of backward propagation of Transformer by Pytorch w/Python3.8 based on the paper: Attention is all you need.

Requirements

torch==2.2.0
numpy==1.22.3

Project Structure

.
|-- Back_Propagation
|   |-- Encoder.py
|   |-- FFN.py
|   |-- LayerNorm.py
|   |-- MultiHead.py
|   |-- __pycache__
|   |   |-- FFN.cpython-38.pyc
|   |   |-- LayerNorm.cpython-38.pyc
|   |   `-- MultiHead.cpython-38.pyc
|   |-- basic_layer.py
|   `-- requirements.txt
|-- LICENSE
|-- README.md

You can test the implementation of back propagation by running the commented code at the bottom of each python file.

Finished

• FFN Layer
• Linear Layer
• Multi-head Attention Layer
• Encoder Layer

To Do

• Embedding Layer
• Decoder Layer
• Encoder
• Decoder
• Transformer

Notice: Without considering dropout

References

• Deep Learning from Scratch(深度学习入门：基于Python的理论与实现)
• Natural Language Processing(深度学习进阶：自然语言处理)
• Pytorch Documentation
• Tensorflow Transformer Tutorial