HRMlaptop

A laptop version of Hierarchical Reasoning Model

HRM Training on Windows Laptop (CPU-Only)

Train a Hierarchical Reasoning Model on your Windows laptop using only CPU - no GPU required!

Quick Start

Prerequisites

• Python 3.12 (via Anaconda or python.org)
• Windows 10/11 (64-bit)
• 8GB RAM minimum (16GB recommended)

Installation

1. Clone or download this project

git clone https://github.com/alessoh/HRMlaptop
cd HRMlaptop

2. Create conda environment (if using Anaconda)

conda create -n HRMlaptop python=3.12
conda activate HRMlaptop

3. Install PyTorch CPU version

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

4. Install remaining dependencies

python quick_install.py

Or manually:

pip install matplotlib psutil tqdm pandas scikit-learn seaborn plotly

Running the Training

Simplest option - for beginners:

python train_simple.py

Full training with all features:

python hrm_trainer.py

Monitor system resources (in separate terminal):

python monitor.py

What's Included

Core Training Scripts

• hrm_trainer.py - Main training script with full features
• train_simple.py - Simplified training for beginners (recommended for first run)
• test_model.py - Test your trained model
• monitor.py - Monitor CPU and memory usage during training
• visualize.py - Create graphs and visualizations of results

Setup Files

• quick_install.py - Install remaining dependencies after PyTorch
• requirements.txt - Package list (reference only)
• .env - Configuration settings
• .gitignore - Git ignore patterns

Installation Helpers

• install.py - Full installation script
• setup_windows.bat - Windows batch installer

Expected Performance

Training Time (10 epochs on MNIST)

CPU Type	Training Time	Accuracy
Intel i5 (8th gen)	90-120 min	~97.5%
Intel i7 (10th gen)	45-60 min	~97.8%
AMD Ryzen 7	40-55 min	~97.8%

Resource Usage

• RAM: 1.5-2.5 GB
• CPU: 60-80% utilization
• Disk: ~500 MB (including dataset)
• Model Size: ~350 KB

Python 3.12 Compatibility Notes

This project is optimized for Python 3.12 with:

• PyTorch 2.8.0+cpu (latest CPU version)
• NumPy 2.1.2 (included with PyTorch)
• All dependencies tested with Python 3.12

Training Output Example

============================================================
HRM Training on Windows CPU
============================================================

System Information:
  CPU: 6 cores, 12 threads
  RAM: 16.0 GB
  Python: 3.12.0
  PyTorch: 2.8.0+cpu
============================================================

Starting training for 10 epochs
Epoch 1/10 [0%] Loss: 0.4521
Epoch 1/10 [20%] Loss: 0.3421
Epoch 1/10 [40%] Loss: 0.2341
...
Epoch 10/10 completed in 324.5s
Train Loss: 0.0234, Train Acc: 98.45%
Val Loss: 0.0412, Val Acc: 97.82%

TRAINING COMPLETE!
Final Validation Accuracy: 97.82%
Total Training Time: 54.3 minutes

Troubleshooting

Issue: "No module named torch"

Solution: Install PyTorch CPU version:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

Issue: NumPy compatibility error

Solution: Use the NumPy that comes with PyTorch (2.1.2). Don't install numpy==1.24.3

Issue: Training is very slow

Solutions:

• Close Chrome and other heavy applications
• Set Windows to High Performance mode
• Ensure good laptop cooling
• Reduce batch size in .env file

Issue: ImportError with Python 3.12

Solution: Use the provided quick_install.py script which installs compatible versions

Issue: High CPU temperature

Solutions:

• Elevate laptop for better airflow
• Use a cooling pad
• Reduce batch size to 16 or 8
• Take breaks between training epochs

Configuration Options

Edit .env file to customize:

# Training parameters
BATCH_SIZE=32          # Reduce to 16 if low on RAM
LEARNING_RATE=0.001    
NUM_EPOCHS=10          # Increase for better accuracy
HIDDEN_DIM=64          # Model size (32 for faster, 128 for better)
NUM_LAYERS=2           # Depth of reasoning layers

# CPU optimization
NUM_THREADS=4          # Set to your physical core count

Project Structure

HRMlaptop/
├── hrm_trainer.py       # Main training script
├── train_simple.py      # Beginner-friendly trainer
├── test_model.py        # Model testing
├── monitor.py           # System monitoring
├── visualize.py         # Results visualization
├── quick_install.py     # Dependency installer
├── README.md           # This file
├── .env                # Configuration
├── requirements.txt    # Package reference
├── data/              # MNIST dataset (auto-created)
├── checkpoints/       # Saved models (auto-created)
└── logs/              # Training logs (auto-created)

Tips for Faster Training

1. Windows Performance Mode

   • Settings → System → Power & battery → Best performance

2. Close Background Apps

   • Especially browsers (Chrome uses lots of RAM)
   • Disable Windows updates during training

3. Optimal Settings for Speed

   BATCH_SIZE=64      # If you have 16GB+ RAM
   HIDDEN_DIM=32      # Smaller model
   NUM_LAYERS=1       # Fewer layers

4. Optimal Settings for Accuracy

   BATCH_SIZE=32      # Default
   HIDDEN_DIM=64      # Default
   NUM_LAYERS=2       # Default
   NUM_EPOCHS=20      # More training

After Training

1. Test your model

   python test_model.py

2. Visualize results

   python visualize.py

3. View training metrics

   • Check training_results.json for detailed metrics
   • View training.log for full training history
   • Open generated .png files for graphs

Understanding the Model

The HRM (Hierarchical Reasoning Model) is optimized for CPU training:

• Small architecture: 64 hidden units (vs 128+ for GPU)
• Efficient layers: 2 reasoning layers with residual connections
• CPU optimizations: Batch normalization, dropout for stability
• Total parameters: ~85K (very lightweight)

Citation

If you use this code for learning or research:

H. Peter Alesso (2025). Hierarchical Reasoning Model: HRM on Your Laptop. 
Chapter 12: Training on Your Windows Laptop.

Support

• Issues: Create an issue on GitHub
• Logs: Check training.log for detailed information
• System Info: See system_info.json for your configuration

License

MIT License - Free for educational and research use

---

Ready to start? Run python train_simple.py and watch your laptop learn to recognize handwritten digits!