MachineTranslation

• This project is a Work in Progress *

This repo contains code for an seq2seq Gujarati -> English translation model, written in PyTorch, based on a 2020 paper.

Project Files

Git Files

• .gitignore: Python git ignore
• requirements.txt: file containing all necessary packages and imports to run the analysis

Directories

• corpus_files: directory containing training sequences and tokens, created by the authors of the original paper
• images: directory with some visualizations of loss during model training
• test_cases: directory for housing tests

Files

• workflow.py: file containing a high-level workflow, used to run the full analysis end-to-end, from data ingenstion to training and beyond.
• training_loop.py: file defining the training and validation steps for the workflow
• textprocessing.py: file containing functions used to parse and clean the sequences and tokens from the input files
• encoder_decoder.py: file containing functions for the encoder and decoders, with commented annotations
• encoder_decoder_clean.py: file containing a clean version (no extra comments or print statements) of encoder_decoder.py

Model Architechture

To-Do List

• Encoder & Decoder Code
  • write core encoder and decoder functions
• Functionalize Jupyter Notebook code
• Run code on GPU
  • Figure out how to configure to "MPS"
• Attention Mechanism
  • Build core Attention mechanism
  • Ensure that decoder accounts for both attention & LSTM
  • Integrate with workflow.py and training_loop.py
• Beam Search Algorithm
  • Understand beam search
  • Write beam search function in isolation
  • Integrate beam search with existing codebase
• Teacher Forcing
  • Implement teacher forcing during training
• Test Coverage
  • write tests for all functions in encoder_decoder.py