hqde 0.1.10

Hierarchical Quantum-Distributed Ensemble Learning Framework

HQDE - Hierarchical Quantum-Distributed Ensemble Learning

A production-ready framework for distributed ensemble learning with quantum-inspired algorithms and adaptive quantization.

HQDE combines quantum-inspired algorithms with distributed computing to deliver superior machine learning performance with significantly reduced memory usage and training time.

Table of Contents

• Key Features
• Installation
• Quick Start
• Architecture Overview
• Quantum-Inspired Algorithms
• Distributed Computing
• Adaptive Quantization
• Configuration
• API Reference
• Performance Benchmarks
• Documentation

---

Key Features

Feature	Description
Up to 17x Faster Training	Ray-based stateful actors with zero-copy data sharing
4x Memory Reduction	Adaptive 4-16 bit quantization based on weight importance
FedAvg Weight Aggregation	Workers share knowledge after each epoch for better accuracy
Ensemble Diversity	Different learning rates and dropout per worker
Production-Ready	Byzantine fault tolerance and dynamic load balancing
Quantum-Inspired	Superposition aggregation, entanglement simulation, QUBO optimization
Distributed	Ray-based MapReduce with O(log n) hierarchical aggregation

---

Installation

From PyPI (Recommended)

pip install hqde

From Source

git clone https://github.com/Prathmesh333/HQDE-PyPI.git
cd HQDE-PyPI
pip install -e .

---

Quick Start

from hqde import SmallImageResNet18, create_hqde_system, make_cifar_training_config

training_config = make_cifar_training_config(
    ensemble_mode='independent',
    batch_assignment='replicate',
    prediction_aggregation='mean',
)

hqde_system = create_hqde_system(
    model_class=SmallImageResNet18,
    model_kwargs={'num_classes': 10},
    num_workers=4,
    training_config=training_config,
)

# Train the ensemble and collect validation metrics each epoch
metrics = hqde_system.train(train_loader, num_epochs=20, validation_loader=test_loader)

# Make predictions (ensemble voting)
predictions = hqde_system.predict(test_loader)

# Evaluate the ensemble directly
eval_metrics = hqde_system.evaluate(test_loader)

# Cleanup resources
hqde_system.cleanup()

Legacy fedavg-style output example:

Epoch 1/40, Average Loss: 2.3045, LR: 0.001000
  → Weights aggregated and synchronized at epoch 1  
Epoch 2/40, Average Loss: 1.8234, LR: 0.000998
  → Weights aggregated and synchronized at epoch 2  

Examples:

python examples/quick_start.py           # Quick demo
python examples/cifar10_synthetic_test.py # CIFAR-10 benchmark
python examples/cifar10_test.py          # Real CIFAR-10 dataset

Current releases log epoch loss, accuracy, and learning rate directly. In independent + replicate mode there is no epoch-end synchronization message because workers stay diverse until inference time.

Training Modes

Use training_config to choose the training behavior that matches your workload:

# True ensemble: preserve diversity, aggregate only at inference
training_config = {
    'ensemble_mode': 'independent',
    'batch_assignment': 'replicate',
}

# Epoch-wise FedAvg/local-SGD style training
training_config = {
    'ensemble_mode': 'fedavg',
    'batch_assignment': 'split',
}

batch_assignment='split' is not PyTorch DDP. Each worker trains locally during the epoch, and weights are averaged only at the epoch boundary when ensemble_mode='fedavg'.

Training Config Notes

training_config = {
    'ensemble_mode': 'independent',
    'batch_assignment': 'replicate',
    'optimizer': 'sgd',
    'learning_rate': 0.1,
    'weight_decay': 5e-4,
    'use_amp': True,
    'label_smoothing': 0.1,
    'warmup_epochs': 5,
    'warmup_start_factor': 0.2,
    'compile_model': False,
    'compile_mode': 'default',
    'prediction_aggregation': 'mean',
}

• Use independent + replicate for true ensemble training.
• Use fedavg + split for epoch-wise averaging with lower memory pressure.
• use_amp activates mixed precision only on CUDA devices.
• quantization_config is only applied in fedavg mode during weight aggregation.

---

Architecture Overview

                    HQDE SYSTEM ARCHITECTURE                      

                                                                  
             
     QUANTUM          DISTRIBUTED           ADAPTIVE       
    INSPIRED       ENSEMBLE       QUANTIZATION     
   ALGORITHMS          LEARNING                            
             
                                                                  

Project Structure

hqde/
 core/
    hqde_system.py           # Main system, workers, quantization
 quantum/
    quantum_aggregator.py    # Superposition and entanglement
    quantum_noise.py         # Quantum noise generation
    quantum_optimization.py  # QUBO and quantum annealing
 distributed/
    mapreduce_ensemble.py      # MapReduce pattern
    hierarchical_aggregator.py # Tree aggregation
    fault_tolerance.py         # Byzantine fault tolerance
    load_balancer.py           # Dynamic load balancing
 utils/
     performance_monitor.py     # System monitoring

---

Quantum-Inspired Algorithms

Note: HQDE uses quantum-inspired algorithms on classical hardware, not actual quantum computers.

Quantum Superposition Aggregation

Combines ensemble predictions using quantum amplitude-like weights:

# Confidence scores converted to quantum amplitudes
amplitudes = sqrt(softmax(confidence_scores))

# Superposition combination
superposition = sum(amplitude_i * prediction_i)

Location: hqde/quantum/quantum_aggregator.py

Entanglement-Based Correlation

Models correlations between ensemble members using an entanglement matrix:

# Symmetric entanglement matrix
entanglement_matrix[i,j] = correlation(model_i, model_j) * strength

# Weight models by their entanglement with others
entangled_weights = softmax(cosine_similarity @ entanglement_matrix)

Location: hqde/quantum/quantum_aggregator.py

Quantum Annealing Optimization

Uses QUBO (Quadratic Unconstrained Binary Optimization) for ensemble selection:

# QUBO formulation for selecting optimal models
qubo_matrix = formulate_qubo(candidate_models, constraints)

# Solve using simulated quantum annealing
solution = quantum_annealing_solve(qubo_matrix)

Location: hqde/quantum/quantum_optimization.py

---

Distributed Computing

HQDE uses Ray for distributed computing with several patterns:

Ray Worker Architecture

# GPUs are automatically divided among workers
@ray.remote(num_gpus=gpu_per_worker)
class EnsembleWorker:
    def train_step(self, data_batch, targets):
        # Each worker trains its own model copy
        ...

MapReduce Weight Aggregation

MAP      →    SHUFFLE    →    REDUCE
Workers       Group by        Aggregate
weights       parameter       weights
              name

Location: hqde/distributed/mapreduce_ensemble.py

Hierarchical Tree Aggregation

Communication Complexity: O(log n)

Level 0 (Root):           [AGG]
                         /     \
Level 1:            [AGG]       [AGG]
                   /    \       /    \
Level 2:        [W1]  [W2]   [W3]  [W4]

Location: hqde/distributed/hierarchical_aggregator.py

Byzantine Fault Tolerance

Tolerates up to 33% faulty or malicious workers:

• Outlier Detection: Median Absolute Deviation (MAD)
• Robust Aggregation: Geometric median
• Reliability Tracking: Source reputation scores

Location: hqde/distributed/fault_tolerance.py

Dynamic Load Balancing

Multi-factor node selection scoring:

• 40% success rate
• 30% current load
• 20% execution speed
• 10% capability match

Location: hqde/distributed/load_balancer.py

---

Adaptive Quantization

Dynamically adjusts precision based on weight importance:

Weight Importance	Bits	Compression
High (critical)	16	2x
Medium (default)	8	4x
Low (redundant)	4	8x

Importance Score = 70% × |weight| + 30% × |gradient|

quantization_config = {
    'base_bits': 8,   # Default precision
    'min_bits': 4,    # High compression for unimportant weights
    'max_bits': 16    # High precision for critical weights
}

Location: hqde/core/hqde_system.py

---

Configuration

Full Configuration Example

from hqde import create_hqde_system

# Quantization settings
quantization_config = {
    'base_bits': 8,
    'min_bits': 4,
    'max_bits': 16
}

# Quantum aggregation settings
aggregation_config = {
    'noise_scale': 0.005,
    'exploration_factor': 0.1,
    'entanglement_strength': 0.1
}

# Create system
hqde_system = create_hqde_system(
    model_class=YourModel,
    model_kwargs={'num_classes': 10},
    num_workers=8,
    quantization_config=quantization_config,
    aggregation_config=aggregation_config
)

---

API Reference

Core Classes

Class	Description	Location
HQDESystem	Main entry point	hqde/core/hqde_system.py
DistributedEnsembleManager	Manages Ray workers	hqde/core/hqde_system.py
AdaptiveQuantizer	Weight compression	hqde/core/hqde_system.py

Quantum Classes

Class	Description	Location
QuantumEnsembleAggregator	Superposition/entanglement aggregation	hqde/quantum/quantum_aggregator.py
QuantumNoiseGenerator	Exploration noise	hqde/quantum/quantum_noise.py
QuantumEnsembleOptimizer	QUBO-based selection	hqde/quantum/quantum_optimization.py

Distributed Classes

Class	Description	Location
MapReduceEnsembleManager	MapReduce pattern	hqde/distributed/mapreduce_ensemble.py
HierarchicalAggregator	Tree aggregation	hqde/distributed/hierarchical_aggregator.py
ByzantineFaultTolerantAggregator	Fault tolerance	hqde/distributed/fault_tolerance.py
DynamicLoadBalancer	Work distribution	hqde/distributed/load_balancer.py

Factory Function

def create_hqde_system(
    model_class,           # PyTorch model class
    model_kwargs,          # Model initialization parameters
    num_workers=4,         # Number of distributed workers
    quantization_config=None,
    aggregation_config=None
) -> HQDESystem

---

Performance Benchmarks

Metric	Traditional Ensemble	HQDE	Improvement
Memory Usage	2.4 GB	0.6 GB	4x reduction
Training Time	45 min	12 min	3.75x faster
Communication	800 MB	100 MB	8x less data
Test Accuracy	91.2%	93.7%	+2.5%

---

Documentation

• HOW_TO_RUN.md - Detailed setup and usage guide
• docs/ - Technical documentation
• examples/ - Working code examples

---

Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/new-feature)
3. Commit your changes (git commit -m 'Add new feature')
4. Push to the branch (git push origin feature/new-feature)
5. Open a Pull Request

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Citation

@software{hqde2025,
  title={HQDE: Hierarchical Quantum-Distributed Ensemble Learning},
  author={Prathamesh Nikam},
  year={2025},
  url={https://github.com/Prathmesh333/HQDE-PyPI}
}

---

Support

• Bug Reports: Create an issue
• Feature Requests: Create an issue
• Questions: Start a discussion