dnaty 1.0.1

Evolutionary Neural Architecture Search — compress any PyTorch model with one function call

dNATY

Evolutionary AI Model Compression

46.5% fewer FLOPs • 1.6× faster inference • 98.85% accuracy retained

Automated model compression using multi-objective evolutionary search. Zero manual engineering. One function call: compress(model, dataset).

---

Why dNATY?

Problem: Most production models are oversized — too slow for real-time inference, too expensive to run.

Existing solutions require:

• Manual architecture tuning
• Days of hyperparameter search
• Accuracy/speed trade-offs

dNATY solves this with episodic memory-guided evolutionary search — operators that worked before are tried more often. Finds Pareto-optimal solutions in minutes.

Proven Results (CIFAR-100)

Model	FLOPs Reduction	Speedup	Accuracy	Time
ResNet-50	-46.5%	1.6×	98.85%	5min
EfficientNet-B0	-40%	1.4×	98.85%	6min
MobileNetV3-Large	-98%	1.8×	97.2%	4min

---

Project Structure

dNATY/
├── dnaty/              # Core compression framework
├── dnaty_saas/         # Production API (FastAPI)
├── frontend/           # Web UI (React + TypeScript)
├── notebooks/          # Experiments & benchmarks
├── scripts/            # Demo & utilities
├── tests/              # Unit tests
└── pyproject.toml      # Package config

---

Quick Start

pip install dnaty

from dnaty import compress
from dnaty.experiments.fast_dataset import FastDataset

# Your existing model (any PyTorch model with Linear layers)
model = your_trained_model

# Your data
ds = FastDataset("MNIST", device="cpu", train_subset=10_000)

# Compress
result = compress(model, ds, target_flops=0.5, n_generations=30)

print(result.summary())
# CompressResult | arch=[128, 64] | FLOPs -46.5% (327680 -> 175104) |
#   params -52.3% (328K -> 156K) | acc=0.9821

---

How It Works

dNATY runs a population of candidate architectures through an evolution loop:

1. Mutate — apply structural operators (add/remove neurons, merge layers, etc.)
2. Train — locally train each candidate for a few epochs
3. Select — NSGA-II Pareto selection: maximize accuracy, minimize FLOPs
4. Remember — episodic memory records which operators helped most; they get picked more often next round

The memory mechanism is dNATY's core innovation. Over generations, the search becomes smarter — not random.

---

API

compress(model, train_data, **kwargs) -> CompressResult

Parameter	Default	Description
model	required	Any nn.Module with Linear layers
train_data	required	DataLoader or FastDataset
target_flops	0.5	Target fraction of original FLOPs (0.5 = 50% less)
n_generations	30	Evolutionary generations
n_pop	15	Population size
device	auto	'cpu' or 'cuda'
seed	None	Fix for reproducibility

CompressResult

result.model              # compressed nn.Module, ready to use
result.accuracy           # validation accuracy
result.flops_reduction    # e.g. 0.465 = 46.5% fewer FLOPs
result.flops_reduction_pct  # same as percentage
result.params_reduction_pct
result.arch               # hidden layer sizes found  [128, 64]
result.summary()          # one-line human-readable summary

---

SaaS API

dNATY ships with a production-ready FastAPI backend.

cd dnaty_saas
cp .env.example .env   # fill DATABASE_URL, JWT_SECRET, ANTHROPIC_API_KEY
uvicorn main:app --reload

POST /api/v1/compress

{
  "description": "classifica defeitos em pecas, precisa rodar no Raspberry Pi",
  "dataset": "MNIST",
  "target_flops": 0.5,
  "n_generations": 30
}

Response 202:

{ "job_id": "a3f2c1b0", "status": "queued", "message": "..." }

GET /api/v1/compress/{job_id}

{
  "status": "completed",
  "result": {
    "accuracy": 0.9821,
    "flops_reduction": 0.465,
    "arch": [128, 64],
    "explanation": "...",        // Claude-generated explanation
    "deployment_code": "..."     // ready-to-use Python code
  }
}

Set ANTHROPIC_API_KEY in .env to enable Claude explanations.
Without it, the endpoint still works — returns template text instead.

---

Getting Started

Web UI

cd frontend
npm install
npm run dev

Open http://localhost:5173

Jupyter Notebooks

• CIFAR-100 Baseline — notebooks/colab_cifar100_notebook.ipynb
• ImageNet Benchmark — notebooks/colab_imagenet_simple_FIXED.ipynb
• CPU Latency — notebooks/benchmark_cpu_latency.py

CLI Demo

python scripts/demo_compress.py           # 20 gens, MNIST (~5 min CPU)
python scripts/demo_compress.py --full    # 30 gens, more accurate
python scripts/demo_compress.py --dataset FashionMNIST

---

Benchmarks

Metric	Value
FLOPs reduction vs. initial arch	-46.5%
FLOPs reduction vs. RandomNAS	better in Pareto front
Speedup to target accuracy	1.6x fewer generations
CL: BWT vs. EWC	6.9x less forgetting

All numbers reproducible with python scripts/prove_it.py.

---

License

BSL 1.1 — free for non-commercial use; contact [e-mail] for commercial licensing.