LSTM_Translator

A neural machine translation system that translates French sentences into English using a Seq2Seq architecture with Bahdanau Attention. This project demonstrates a complete pipeline from preprocessing and embedding to training, evaluation, and visualization through a Pygame interface.

---

🚀 Overview

This project implements a French-to-English machine translation model using a Bidirectional LSTM Encoder, Bahdanau Attention, and an LSTM Decoder. It leverages pretrained FastText word embeddings (cc.fr.300.vec.gz and cc.en.300.vec.gz) to provide meaningful semantic initialization and better generalization.

---

📁 Project Structure

LSTM_Machine_Translation/
│
├── dataset_class.py # Builds vocab and embedding matrix from FastText
├── model.py # Encoder-Decoder with Bahdanau Attention
├── testing_single_data.py # Prepares training data using vocab and embeddings
├── testing_data.py # Prepares testing data in same format
├── app.py # Pygame app to visualize / test translation, includes attention heatmap display
├── README.md # This file
└── .gitignore # Ignores large .vec.gz files

🧠 Model Architecture

🔹 Encoder

• A Bidirectional LSTM is used to capture both forward and backward context of the French input sentence.
• The hidden states from both directions are concatenated and passed to the decoder.

🔹 Bahdanau Attention

• Instead of just relying on the final encoder state, Bahdanau Attention (also known as additive attention) computes a context vector at each decoding step by attending to all encoder hidden states.
• This allows the decoder to dynamically focus on relevant parts of the input sentence during translation.
• It improves translation quality especially for longer or more complex input sentences.

🔹 Decoder

• The decoder is an LSTM that generates one word at a time.
• It takes in the previous word, previous hidden state, and the attention-generated context vector to produce the next word.
• The decoder is trained using teacher forcing for stability.

---

📦 FastText Embeddings

We use:

• cc.fr.300.vec.gz for French word vectors
• cc.en.300.vec.gz for English word vectors
• Due to big sizes couldn't be added here.

---

🧾 Data Preparation

• dataset_class.py:
  Reads text data, tokenizes, builds vocab, and loads FastText vectors to create the embedding matrix.

• testing_single_dataset.py:
  Uses the vocab and embedding matrix to tokenize, numericalize, pad, and batch training data.

• testing_data.py:
  Prepares testing samples similarly for inference/evaluation.

---

🎮 Pygame Interface (app.py)

A simple Pygame-based GUI to type a French sentence and see the English translation predicted by the model. It also displays a visual attention heatmap, showing which parts of the input the model focuses on at each step—making the project both interactive and interpretable.

---

🛠 Training (Included in main.py)

• Use dataset_class.py for loading training batches, embedding matrix, and vocabulary
• Pass through model.py’s encoder-decoder
• Use cross-entropy loss and Adam optimizer

---

Notes

This uses Python 3.11 as the newer versions aren't compatible with CUDA Torch needed for GPU usage while training Checkpoints file saves the models every 5 epoches and the best model yet For training, The model requires two embeddings French (cc.fr.300.vec.gz) and English (cc.en.300.vec.gz). These are fasttext embeddings any other pretrained can be used as well

• French: https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.fr.300.vec.gz
• English: https://dl.fbaipublicfiles.com/fasttext/vectors-crawl/cc.en.300.vec.gz