Language Translation (Encoder-Decoder)

This project implements a Seq2Seq language translation model using an LSTM-based encoder-decoder in PyTorch.

Supported examples:

• English -> French
• English -> Tamil

Project Structure

• train_seq2seq.py - training, evaluation, and interactive translation
• app.py - Flask frontend for browser-based translation
• templates/index.html - main frontend page
• static/styles.css - frontend styling
• requirements.txt - Python dependencies
• data/sample_en_fr.tsv - tiny English-French sample dataset
• data/sample_en_ta.tsv - tiny English-Tamil sample dataset

Dataset Format

Use a tab-separated file with one sentence pair per line:

english sentence<TAB>target sentence

Setup

pip install -r requirements.txt

Train

English -> French:

python train_seq2seq.py --data_path data/sample_en_fr.tsv --source_lang en --target_lang fr --epochs 300

Train on a larger French dataset file such as C:/Users/admin/OneDrive/Desktop/fra.txt:

python train_seq2seq.py --data_path "C:/Users/admin/OneDrive/Desktop/fra.txt" --source_lang en --target_lang fr --max_samples 50000 --epochs 15 --batch_size 64

Force GPU training when CUDA is available:

python train_seq2seq.py --data_path "C:/Users/admin/OneDrive/Desktop/fra.txt" --source_lang en --target_lang fr --max_samples 20000 --epochs 8 --batch_size 64 --device cuda

English -> Tamil:

python train_seq2seq.py --data_path data/sample_en_ta.tsv --source_lang en --target_lang ta --epochs 300

Translate

After training, the script saves:

• artifacts/config.json
• artifacts/model.pt
• artifacts/vocab.json

Run interactive inference:

python train_seq2seq.py --mode translate

Frontend

Start the web app:

python app.py

Then open:

http://127.0.0.1:5000

The frontend uses the latest files in artifacts/. If those files are missing, train the model first.

Large Dataset Tips

• fra.txt contains extra metadata columns, and the trainer automatically uses only the first two columns.
• Start with --max_samples 50000 for a manageable first run.
• After training on fra.txt, restart the Flask app so it picks up the new artifacts.
• Use --device cuda to force GPU training, --device cpu to force CPU, or leave the default auto.

Notes

• The included datasets are intentionally tiny, so translations are only for demonstration.
• For real performance, replace the sample dataset with a larger parallel corpus.
• The model uses greedy decoding for simplicity.