An independent, lightweight Deep Learning project featuring a 2-Layer LSTM network built entirely in PyTorch (completely free of TensorFlow dependencies). The model is uniquely trained on a blended stream of contrasting text corpora—classic Shakespearean theatrical drama and modern Elasticsearch technical documentation—allowing it to predict the next sequential word across distinct writing styles.
Below is a demonstration of the interactive Gradio interface processing inputs, handling out-of-vocabulary terms gracefully, and executing next-word tensor inferences in real-time.
demo.mp4
💡 Note: If you are editing this README on GitHub, drag and drop your renamed
demo.mp4video directly into the markdown editor to generate your unique cloud asset stream link, then paste it into thesrc=""attribute above.
Standard feed-forward neural networks fail to process contextual language because they lack sequential memory. This architecture overcomes that by leveraging recurrent state-tracking:
- Pure Python Tokenization: Implements a custom text preprocessor to build vocabulary frequency indexes and handle left-aligned pre-padding sequences natively without framework overhead.
-
Embedding Layer: Transforms text tokens into dense continuous vectors (
$128$ -dimensional space) capturing semantic relationships. -
Stacked LSTM Layers: Utilizes a deeper
$2$ -layer LSTM network with$256$ hidden dimensions to track complex long-range textual context across varied domains. -
Regularization (Dropout): Integrates a
dropout=0.3rate between sequential layers to prevent overfitting on specific dataset biases. - Gradio Web Interface: Deploys a clean GUI complete with automated runtime example inputs for streamlined evaluation.
word-predictor/
├── assets/
│ └── demo.mp4 # Professional demo video recording
├── app.py # Main PyTorch & Gradio application script
├── requirements.txt # Environment dependencies list
└── README.md # Project documentation