Seq2Seq Neural Machine Translation

This repository contains a Sequence-to-Sequence (Seq2Seq) neural machine translation model for translating from French to English, along with inference utilities.

Files

• seq2seq.ipynb - Main training notebook for the Seq2Seq model with attention mechanism
• inference.ipynb - Inference notebook for generating translations using trained models
• fra.csv - Test Data for inference
• tatoeba/eng-fra.txt - Train Data from Tatoeba with english and french sentences

Project Overview

This project implements a neural machine translation system using:

• Encoder-Decoder Architecture with attention mechanism
• Bidirectional GRU encoder
• GRU-based Decoder with attention

The model is trained on French-English parallel text data and supports both training and inference workflows.

Requirements

• Python 3.7+
• PyTorch
• pandas
• tqdm

Install dependencies:

pip install torch pandas tqdm

Usage

Training (seq2seq.ipynb)

1. Prepare training data in TSV format (source \t target)

2. Configure hyperparameters:

   • Embedding dimension
   • Hidden dimension
   • Number of layers
   • Dropout rate
   • Learning rate
   • Batch size

3. Run the notebook cells to:

   • Load and preprocess data
   • Build vocabularies
   • Initialize and train the model
   • Save trained weights

Inference (inference.ipynb)

1. Load pre-trained model weights and vocabularies
2. Provide input sentences for translation
3. Generate translations using beam search or greedy decoding

Example:

input_sentence = "Comment allez-vous?"
translation = translate(input_sentence)
print(translation)  # "How are you?"

Data Format

Training data should be in TSV format:

French sentence	English sentence
Bonjour	Hello
Comment allez-vous?	How are you?

Model Architecture

Encoder

• Embedding layer
• Bidirectional GRU (2 layers)
• Linear projection layer

Decoder

• Embedding layer
• GRU with attention
• Output projection layer

Attention Mechanism

• Luong-style attention
• Scores computed over encoder outputs