Skip to content

QUICKSTART.md

Latest commit

 

History

History
241 lines (194 loc) · 6.14 KB

QUICKSTART.md

File metadata and controls

241 lines (194 loc) · 6.14 KB
Raw

Quick Start Guide

This document provides a quick overview of how to get started with this deep learning project.

Prerequisites

  • Python 3.8 or higher
  • 2GB free disk space (for dataset download)
  • 4GB RAM (recommended)

Installation Steps

1. Install Dependencies

pip install -r requirements.txt

2. Verify Installation

python verify_setup.py

You should see:

✓ Python Version: 3.x.x
✓ Checking dependencies:
  ✓ NumPy - OK
  ✓ Matplotlib - OK
  ✓ Scikit-learn - OK
✓ Checking project modules:
  ✓ utils - OK
  ✓ data_loader - OK
  ✓ mlp - OK
  ✓ autoencoder - OK
  ✓ rbm - OK
ALL CHECKS PASSED!

Running the Project

Option 1: Train All Models (Recommended for First Run)

python train.py

This will:

  • Download Fashion-MNIST dataset (∼220MB)
  • Train MLP for classification (∼2-3 minutes)
  • Train Dense Autoencoder (∼2-3 minutes)
  • Train RBM (∼3-4 minutes)
  • Generate visualizations
  • Total time: ∼10-15 minutes on CPU

Output files created in results/ directory:

  • training_history.png - Training curves for all models
  • reconstructions.png - Autoencoder reconstruction examples
  • rbm_filters.png - Learned RBM filters

Option 2: Detailed Evaluation

python evaluate.py

This will:

  • Run MLP evaluation with confusion matrix
  • Analyze Autoencoder reconstruction error
  • Analyze RBM feature learning
  • Generate detailed performance reports

Additional output files:

  • mlp_confusion_matrix.png - Classification confusion matrix
  • autoencoder_analysis.png - Reconstruction error distribution
  • rbm_analysis.png - Feature activation analysis

Expected Performance

After running train.py, you should see:

[2/5] Training Multilayer Perceptron (MLP)...
Epoch 5/30 - Loss: 0.2543, Acc: 92.45% - Val Loss: 0.2651, Val Acc: 91.23%
Epoch 10/30 - Loss: 0.1876, Acc: 94.12% - Val Loss: 0.2145, Val Acc: 93.45%
...
Epoch 30/30 - Loss: 0.0987, Acc: 96.78% - Val Loss: 0.1654, Val Acc: 94.23%
  ✓ MLP Test Accuracy: 88.45%

[3/5] Training Dense Autoencoder...
Epoch 5/30 - Train Loss: 0.001234, Val Loss: 0.001567
...
  ✓ Autoencoder Test MSE: 0.000891

[4/5] Training Restricted Boltzmann Machine (RBM)...
Epoch 5/30 - Free Energy: -123.45
...
  ✓ RBM Test Reconstruction Error: 0.145234

Project Structure

d:\MLP(Leslie)\
├── utils.py                 # Core utility functions
├── data_loader.py          # Fashion-MNIST loader
├── mlp.py                  # MLP implementation
├── autoencoder.py          # Autoencoder implementation
├── rbm.py                  # RBM implementation
├── train.py                # Main training script
├── evaluate.py             # Evaluation script
├── verify_setup.py         # Verification script
├── requirements.txt        # Python dependencies
├── README.md              # Full documentation
├── QUICKSTART.md          # This file
└── results/               # Output directory (created after first run)
    ├── training_history.png
    ├── reconstructions.png
    ├── rbm_filters.png
    ├── mlp_confusion_matrix.png
    ├── autoencoder_analysis.png
    └── rbm_analysis.png

Customization

Modify Hyperparameters

Edit train.py to adjust:

# MLP Configuration (around line 63)
mlp = MLP(
    input_size=784,
    hidden_size=256,  # Change from 128
    output_size=10,
    hidden_activation='relu',
    learning_rate=0.01  # Change from 0.05
)

# Autoencoder Configuration (around line 88)
autoencoder = DenseAutoencoder(
    input_size=784,
    hidden_sizes=[512, 256],  # Change from [256]
    latent_size=32,
    learning_rate=0.001,  # Change as needed
    activation='relu'
)

# RBM Configuration (around line 112)
rbm = RBM(
    n_visible=784,
    n_hidden=256,  # Change from 100
    learning_rate=0.001
)

# Training parameters (update all sections)
history_mlp = mlp.train(
    X_train, y_train_onehot,
    epochs=50,  # Change from 30
    batch_size=32,  # Change from 64
)

Experiment with Different Architectures

MLP:

# Try different hidden layer sizes
hidden_size = 256  # vs 128, 64, 512
hidden_activation = 'sigmoid'  # vs 'relu'
learning_rate = 0.01  # vs 0.05, 0.1

Autoencoder:

# Try different encoder/decoder depths
hidden_sizes = [512, 256, 128]  # vs [256]
latent_size = 16  # vs 32, 64
sparse = True  # Enable sparsity

RBM:

# Try different hidden dimensions
n_hidden = 200  # vs 100, 50
cd_k = 5  # More Gibbs sampling steps

Troubleshooting

Issue: "ModuleNotFoundError: No module named X"

Solution: Run pip install -r requirements.txt again

Issue: "CUDA out of memory"

Solution: Reduce batch_size in train.py (e.g., 32 instead of 64)

Issue: Training is very slow

Solution:

  • This is expected on CPU. Consider GPU if available.
  • Reduce epochs for quicker testing
  • Reduce batch_size for faster iteration

Issue: Dataset download fails

Solution:

Next Steps

  1. ✓ Run verify_setup.py to ensure everything works
  2. ✓ Run train.py to train all models
  3. ✓ Run evaluate.py to see detailed analysis
  4. ✓ Experiment with hyperparameters
  5. ✓ Study the code to understand implementations
  6. ✓ Check visualizations in results/ directory

Further Learning

  • Modify network architectures in mlp.py, autoencoder.py, rbm.py
  • Implement new activation functions in utils.py
  • Add new metrics or loss functions
  • Extend to other datasets
  • Implement CNN or other architectures

References

Support

For issues or questions:

  1. Check README.md for detailed documentation
  2. Review code comments in implementation files
  3. Run evaluate.py to debug training
  4. Experiment with simpler configurations first

Happy Learning! 🚀