Sanskrit-to-English Neural Machine Translation

A character-level Transformer model for translating English text to Sanskrit using PyTorch.

Overview

This project implements a complete neural machine translation system that converts English text to Sanskrit (Devanagari script) using a Transformer architecture with multi-head attention mechanisms.

Features

• Transformer Architecture: Full encoder-decoder with multi-head attention
• Character-Level Translation: Fine-grained tokenization for better accuracy
• Complete Vocabulary: Comprehensive Sanskrit character set

Quick Start

1. Install Dependencies

pip install torch numpy matplotlib jupyter

2. Train the Model (Optional)

cd transformer/
jupyter notebook final_transformer.ipynb
# Run all cells to train from scratch

Model Architecture

• Type: Encoder-Decoder Transformer
• Dimensions: 512 (d_model), 2048 (FFN)
• Attention Heads: 8
• Layers: 1 (configurable)
• Vocabulary: 89 Sanskrit + 183 English characters
• Max Length: 200 characters
• Training: Adam optimizer, Cross-entropy loss

Sample Translations:

• "I am here" → अहम् अत्र अस्मि
• "Do work don't expect result" → कर्म कुर्वन्तु फलं मा प्रत्याशयन्तु

Technical Details

• Framework: PyTorch 2.0+
• Training Data: English-Sanskrit parallel corpus
• Tokenization: Character-level with special tokens
• Attention: Multi-head self-attention + cross-attention
• Masking: Look-ahead and padding masks

Training Process

1. Data Preparation: Filter valid sentence pairs
2. Vocabulary Building: Character-level tokenization
3. Model Initialization: Xavier uniform weights
4. Training Loop: 50 epoch with checkpoint saving
5. Evaluation: Real-time translation testing

Performance

• Training: ~50 epochs on parallel corpus
• Inference: Real-time character generation
• Accuracy: Not accurate/Accurate over large corpus