Deep Learning Assignment - KIIT University

A deep learning project that classifies movie reviews as positive or negative using LSTM (Long Short-Term Memory) neural networks.

• Overview
• Requirements
• Installation
• How to Run
• Project Structure
• Model Architecture
• Results
• Dataset Information
• Assignment Details

This project implements a sentiment analysis model using Bidirectional LSTM networks to classify IMDB movie reviews. The model processes text sequences and predicts whether a review expresses positive or negative sentiment.

Key Features:

• Binary sentiment classification (Positive/Negative)
• Bidirectional LSTM architecture for better context understanding
• Word embeddings for semantic representation
• Comprehensive evaluation metrics
• Visualization of training progress
• Sample predictions on custom reviews

• Python 3.8 or higher

tensorflow>=2.10.0 numpy>=1.21.0 pandas>=1.3.0 matplotlib>=3.4.0 seaborn>=0.11.0 scikit-learn>=1.0.0 jupyter>=1.0.0

• Minimum: 8GB RAM, CPU-based training (~30-45 minutes)
• Recommended: 16GB RAM, GPU support for faster training (~5-10 minutes)

cd /path/to/lstm-project

# Create virtual environment python -m venv venv # Activate virtual environment # On macOS/Linux: source venv/bin/activate # On Windows: venv\Scripts\activate

# Install all dependencies pip install tensorflow numpy pandas matplotlib seaborn scikit-learn jupyter # Or if you have a requirements.txt: pip install -r requirements.txt

python -c "import tensorflow as tf; print('TensorFlow version:', tf.__version__)"

1. Launch Jupyter Notebook:

   jupyter notebook

2. Open the notebook:

   • Navigate to imdb_sentiment_analysis.ipynb
   • Click to open

3. Run the cells:

   • Click "Cell" → "Run All" to execute all cells
   • Or run cells individually with Shift + Enter

4. View outputs:

   • Training progress and metrics will be displayed inline
   • Plots and visualizations will appear below each cell

# Install JupyterLab if not already installed pip install jupyterlab # Launch JupyterLab jupyter lab # Open imdb_sentiment_analysis.ipynb

1. Open VS Code
2. Install the Jupyter extension
3. Open imdb_sentiment_analysis.ipynb
4. Click "Run All" or run cells individually

lstm-project/ │ ├── imdb_sentiment_analysis.ipynb # Main Jupyter notebook with complete implementation ├── README.md # This file - project documentation ├── requirements.txt # Python package dependencies (optional) └── (Generated after running notebook) └── imdb_sentiment_lstm_model.h5 # Saved trained model 

The model uses a Bidirectional LSTM architecture:

1. Embedding Layer (10,000 → 128) - Converts word indices to dense vectors - Vocabulary size: 10,000 most frequent words - Embedding dimension: 128 2. Dropout Layer (20%) - Regularization to prevent overfitting 3. Bidirectional LSTM Layer (128 units) - Processes sequences in both directions - Returns sequences for stacking 4. Dropout Layer (20%) 5. LSTM Layer (64 units) - Second recurrent layer for feature extraction 6. Dropout Layer (20%) 7. Dense Layer (64 units, ReLU) - Fully connected layer 8. Dropout Layer (20%) 9. Output Layer (1 unit, Sigmoid) - Binary classification output 

• Validation Split: 20% of training data
• Early Stopping: Monitors validation loss (patience=3)
• Learning Rate Reduction: Reduces LR on plateau (factor=0.5, patience=2)
• Dropout: 20% dropout rate for regularization

The model typically achieves:

• Test Accuracy: 85-88%
• Precision: 0.85-0.88
• Recall: 0.85-0.88
• F1-Score: 0.85-0.88

1. Dataset Exploration

   • Class distribution visualization
   • Sequence length distribution
   • Sample decoded reviews

2. Training Curves

   • Accuracy over epochs (training & validation)
   • Loss over epochs (training & validation)

3. Evaluation Metrics

   • Confusion matrix heatmap
   • Classification report (precision, recall, F1-score)
   • Detailed performance statistics

4. Sample Predictions

   • Correct predictions with confidence scores
   • Incorrect predictions for error analysis
   • Custom review predictions

• Source: Keras Datasets / TensorFlow Datasets
• Total Samples: 50,000 reviews
• Training Set: 25,000 reviews
• Test Set: 25,000 reviews
• Classes: Binary (Positive/Negative)
• Class Balance: 50% positive, 50% negative

• Average Review Length: ~230 words
• Vocabulary: 88,585 unique words (using top 10,000)
• Format: Pre-tokenized integer sequences
• Language: English

1. Tokenization: Words converted to integer indices
2. Vocabulary Limiting: Top 10,000 most frequent words
3. Padding: Sequences padded/truncated to 250 words
4. Padding Strategy: Post-padding (zeros added at end)

✅ Dataset Loading

• Loaded IMDB dataset via Keras
• Explored dataset statistics and distribution
• Displayed sample reviews

✅ Text Preprocessing

• Tokenized text data (word to integer mapping)
• Padded sequences to uniform length (250 words)
• Created vocabulary of top 10,000 words

✅ Model Building

• Built Bidirectional LSTM model
• Included embedding layer (128 dimensions)
• Used sigmoid activation for binary classification

✅ Training

• Split data into train/validation/test sets
• Compiled with binary cross-entropy loss
• Used Adam optimizer
• Trained for up to 10 epochs with early stopping

✅ Evaluation

• Evaluated on test set
• Reported accuracy, loss, precision, recall, F1-score
• Generated confusion matrix
• Provided sample predictions

• ✅ Jupyter Notebook (.ipynb) with complete implementation
• ✅ Proper comments explaining each step
• ✅ Training/validation curves
• ✅ Evaluation metrics and confusion matrix
• ✅ Sample predictions
• ✅ README file with setup instructions

1. TensorFlow Installation Fails

# Try installing TensorFlow CPU version explicitly pip install tensorflow-cpu

2. Out of Memory Error

# Reduce batch size in the notebook batch_size = 64 # Instead of 128

3. Jupyter Kernel Dies During Training

• Close other applications
• Reduce model complexity or batch size
• Use CPU instead of GPU if GPU memory is limited

4. Missing Word Index

# If word_index fails to load, the notebook handles this automatically # The dataset will still work for training and evaluation

Potential enhancements to improve performance:

1. Pre-trained Embeddings: Use GloVe or Word2Vec embeddings
2. Attention Mechanism: Add attention layers for better focus
3. Ensemble Methods: Combine multiple models
4. Data Augmentation: Synonym replacement, back-translation
5. Hyperparameter Tuning: Grid search or Bayesian optimization
6. Transformer Models: Experiment with BERT or similar architectures

• Dataset: IMDB Dataset - Keras
• TensorFlow Documentation: https://www.tensorflow.org/
• LSTM Paper: Hochreiter & Schmidhuber (1997)
• Sentiment Analysis: Stanford NLP Sentiment

Aniruddha Mukherjee KIIT University Deep Learning Assignment

Submit to: sunil.sawantfcs@kiit.ac.in Files to Submit:

• imdb_sentiment_analysis.ipynb
• README.md

This project is created for educational purposes as part of a university assignment.

• KIIT University Faculty
• TensorFlow/Keras Development Team
• Andrew Maas et al. for the IMDB dataset