nhge 0.1.5.dev0

Neuro-Harmonic Graph Engine — iterative harmonic resonance ML architecture

Neuro-Harmonic Graph Engine (NHGE)

#IDEA INITIATED BY H.A.Mwala (Full name HEKIMA A. MWALA) #TANZANIAN A novel machine learning architecture that replaces transformer parallelism with iterative harmonic resonance over a dynamic graph structure.

---

Key Idea

Transformers compute attention in one parallel pass over all tokens simultaneously. NHGE iterates — tokens are graph nodes that resonate with each other, updating their states step by step until the graph reaches harmonic convergence.

Property	Transformer	NHGE
Processing	Parallel attention	Iterative harmonic resonance
Depth	Fixed layers	Dynamic (stops at convergence)
Edge weights	Attention scores	Similarity × phase coherence
Memory	O(N²) attention matrix	O(N²) but iterative, not stacked
Simple inputs	Same cost as complex	Converges faster → cheaper

---

Architecture

Input tokens
     ↓
[Embedding + positional + phase initialisation]
     ↓
┌─────────────────────────────────────────┐
│  Harmonic iteration loop (max T steps)  │
│                                         │
│  HarmonicEdgeLayer                      │
│    edge_w = softmax(Q·K/√d) × cos(Δθ)  │
│                                         │
│  HarmonicNodeUpdate                     │
│    h_v ← h_v + FFN(h_v + Σ w·h_u)      │
│                                         │
│  PhaseUpdate                            │
│    θ ← θ + α·tanh(W·h)·π               │
│                                         │
│  Convergence: ||h_t − h_{t−1}|| < ε    │
└─────────────────────────────────────────┘
     ↓
[Graph readout: mean / CLS / attention pool]
     ↓
Output logits

---

Files

File	Purpose
nhge_model.py	Core NHGE architecture (all layers + full model)
nhge_tokenizer.py	Word / char / subword tokenizer
nhge_trainer.py	Training engine with warmup LR, AMP, checkpoints
nhge_inference.py	Generation, classification, embeddings
demo.py	Runnable end-to-end demo (no GPU needed)

---

Quick start

from nhge_model     import nhge_small
from nhge_tokenizer import NHGETokenizer
from nhge_inference import NHGEInference

# 1. Build tokenizer
tok = NHGETokenizer(mode="word")
tok.build_vocab(your_texts, min_freq=2)

# 2. Build model
model = nhge_small(vocab_size=tok.vocab_size, num_classes=2)

# 3. Inference
inf = NHGEInference(model, tok, device="cuda")
results = inf.classify(["Your input text here"], label_names=["neg", "pos"])

# 4. Generate
text = inf.generate("The harmonic graph", max_new_tokens=50)

---

Training

from nhge_trainer import NHGETrainer, TokenDataset
from torch.utils.data import DataLoader

dataset = TokenDataset(encoded_tokens, labels=class_labels, max_len=128)
loader  = DataLoader(dataset, batch_size=32, shuffle=True)

trainer = NHGETrainer(model, loader, val_loader=val_loader,
                      task="cls", lr=3e-4, device="cuda")
history = trainer.train(epochs=20)

---

Model sizes

Name	d_model	n_heads	n_layers	max_iter	~Params
nhge_small	128	4	2	6	~3M
nhge_base	512	8	4	8	~85M
nhge_large	1024	16	6	10	~340M

---

Requirements

torch >= 2.0

No other dependencies required for core functionality.

---

Run the demo

python demo.py

---

Theory

The harmonic edge weight combines two signals:

1. Semantic similarity — scaled dot product of Q and K projections
2. Phase coherence — cos(θ_i − θ_j) between learned phase angles

Nodes in phase amplify each other's signals. Out-of-phase nodes cancel. This mimics oscillatory dynamics in biological neural networks, where synchronised firing encodes binding of related concepts.

The phase angles are updated each iteration via a damped gradient:

θ ← θ + α · tanh(W·h) · π

This allows the network to "tune" its resonance frequencies as information propagates — a form of learned synchronisation.

---

Licence

MIT — use freely, extend openly, credit the NHGE project.

MIT License

Copyright (c) 2026 H.A. Mwala

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.