Poem Generator Using LSTMs

This project uses Long Short-Term Memory (LSTM) neural networks to generate poems based on a dataset from the Poetry Foundation. It includes data cleaning, model training, and poem generation with customizable parameters.

---

Prerequisites

• Python version 3.8 - 3.11 must be installed on your system.

---

Project Structure

Poem-Generator-Using-LSTMs/
├── cleaning.py                # Preprocesses raw data using regex & pandas
├── main.py                    # Main script for training & generation
├── requirements.txt           # Project dependencies
├── datos_limpios/             # Cleaned text data files
├── Modelos/                   # Trained models (includes a sample model)
├── Results/
│   ├── Evaluations.txt        # Model summaries & metadata
│   └── Poemas_Generados.txt   # Generated poems and generation settings

• cleaning.py: Cleans the raw poem data using regular expressions and pandas, producing four text files of different character lengths in the datos_limpios/ folder.
• Modelos/: Stores all trained models. A pretrained model is provided so you can generate poems without training one.
• Results/:
  • Evaluations.txt: Contains summaries of each trained model — parameters, dataset info, creation time, etc.
  • Poemas_Generados.txt: Logs every generated poem with the model used, seed phrase, temperature, generation time, and more.

---

Steps to Run the Project

1. Clone the repository and download the dataset

git clone https://github.com/cesarsiuu2316/Poem-Generator-Using-LSTMs.git

Download the dataset PoetryFoundationData.csv and place it into the root folder if not already included.

---

2. Create and Activate a Virtual Environment

It’s recommended to use a virtual environment:

# Create the virtual environment using pip
python -m venv venv

Or specify Python version:

py -3.11 -m venv venv

Activate it

# Activate environment
venv\Scripts\activate

---

3. Install Dependencies

With the virtual environment active:

pip install -r requirements.txt

---

4. Configure the Project

In main.py:

• Modify global variables to set training parameters.
• Adjust paths to the dataset and model file.
• To use the an existing pretrained model in the Modelos/ folder, simply update the path accordingly — training will be skipped.

---

5. Run the Project

# Run to clean data, not required.
python cleaning.py

# Train and run models
python main.py

---

Model Architecture & Performance

The best performing model included in Modelos/ (Model 5) utilizes a deep learning architecture optimized for sequence generation.

Architecture Details

• Embedding Layer: Maps the vocabulary (41 unique characters) to a 128-dimensional dense vector space.
• LSTM Layers: Two stacked LSTM layers with 512 units each. The first LSTM layer returns sequences to feed into the second, capturing long-range dependencies in the text.
• Dropout Layers: Applied after each LSTM layer to prevent overfitting.
• Dense Layer: A final dense layer with a softmax activation function to predict the probability distribution of the next character.
• Total Parameters: ~10.3 Million

Training Results

• Dataset: Kaggle Poetry Foundation dataset (41 unique characters)
• Training Epochs: 20
• Optimizer: Adam (Learning rate: 0.0005)
• Loss Function: Sparse Categorical Crossentropy
• Final Accuracy: ~75.07%