BERT Model for IMDB Movie Review Sentiment Classification

📋 Overview

• Model: BERT-base-uncased (110M parameters)
• Task: Binary sentiment classification (Positive/Negative)
• Dataset: IMDB movie reviews (50,000 reviews)
• Framework: PyTorch + Hugging Face Transformers

🚀 Quick Start

1. Install Dependencies

2. Train the Model

• Download and prepare the IMDB dataset
• Load pre-trained BERT model
• Train for 3 epochs with early stopping
• Save the best model to models/best_model.pt
• Generate training curves in results/

3. Evaluate the Model

• Load the trained model
• Evaluate on the test set
• Generate classification report and confusion matrix

4. Make Predictions

📁 Project Structure

IMDB/
├── config.py           # Configuration and hyperparameters
├── data_loader.py      # Data loading and preprocessing
├── model.py            # BERT classifier architecture
├── train.py            # Training script
├── evaluate.py         # Evaluation script
├── inference.py        # Inference script
├── utils.py            # Utility functions
├── requirements.txt    # Python dependencies
├── models/             # Saved models
│   ├── best_model.pt
│   └── tokenizer/
└── results/            # Training outputs
    ├── training_curves.png
    └── confusion_matrix.png

🔧 Configuration

• MODEL_NAME: "bert-base-uncased"
• BATCH_SIZE: 16
• LEARNING_RATE: 2e-5
• NUM_EPOCHS: 3
• MAX_LENGTH: 512 tokens

📊 Model Architecture

BERTClassifier
├── BERT Base (12 layers, 768 hidden size)
├── Dropout (p=0.3)
└── Linear Classifier (768 → 2)

🎯 Expected Performance

• Training Accuracy: ~95%
• Validation Accuracy: ~90%
• Test Accuracy: ~88-90%

💡 Usage Examples

Training with Custom Config

Making Predictions

📈 Training Process

1. Data Loading: IMDB dataset split into 80% train, 10% validation, 10% test
2. Tokenization: BERT tokenizer with max length 512
3. Optimization: AdamW optimizer with linear warmup
4. Training: 3 epochs with early stopping (patience=2)
5. Evaluation: Accuracy, precision, recall, F1-score

🛠️ Advanced Features

• Early Stopping: Prevents overfitting
• Gradient Clipping: Stabilizes training
• Learning Rate Warmup: Improves convergence
• Checkpoint Saving: Saves best model automatically
• Visualization: Training curves and confusion matrix

📝 Notes

• Training on CPU will be slow (~2-3 hours per epoch)
• GPU recommended for faster training (~10-15 minutes per epoch)
• Model requires ~2GB disk space
• First run will download BERT weights (~440MB)

🤝 Contributing

📄 License