v2.0.1 — Documentation corrections

v2.0.1 — Documentation corrections only

No API, algorithm, or benchmark changes. Fixes wrong numbers published in v2.0.0 README.

Fixed

• Electrical Fault Detect: −86.0% → −83.0% FLOPs; 99.04% → 99.25% accuracy
• Breast Cancer (UCI): 100.0% → 99.56% accuracy
• Headline range: "8–86%" → "8–83% fewer FLOPs"
• METHODOLOGY.md: max dataset corrected (−86% Electrical Fault → −83.4% Dry Bean Quality)

Install

pip install dnaty==2.0.1

v2.0.0 — Hardware-aware Latency NAS, N:M Sparsity, INT8

What's new in v2.0.0

Hardware-aware Latency NAS

• LatencyEvolver — bi-objective NAS minimising (1−accuracy, latency_ms) instead of FLOPs. Uses real ONNX Runtime measurement; optional GBM surrogate for 80% of evaluations.
• compress(model, data, target="latency", hw_target="rpi4") — new target parameter. Fully backwards-compatible.
• detect_hw() / latency_scale(device) / estimate_latency(ms, from, to) — cross-device estimation without hardware. Calibrated: RPi4 (9×), RPi5 (4.5×), Jetson Nano (5×), Apple M1 (0.6×).
• LatencyPredictor — GBM (LightGBM) surrogate. Degrades gracefully to analytical fallback.

N:M Structured Sparsity

• apply_nm_sparsity(model, n=2, m=4) — 2:4 structured sparsity baked into weights, ONNX-ready.
• sparsity_stats(model) — global and per-layer sparsity report.
• compress(model, data, sparsity="2:4") — apply automatically after NAS.

Quantization-aware Search

• QuantAwareEvolver — fitness evaluated under INT8. Filters architectures that degrade under quantization during search. ~20% overhead vs standard.
• compress(model, data, quant_aware=True) — enable via parameter.
• result.quantize() — apply torch.quantization.quantize_dynamic in one call.

Pandas / NumPy Auto-input

• compress() now accepts (X, y) tuple, pandas.DataFrame, or numpy.ndarray directly — no manual DataLoader needed.

Telemetry Flywheel

• Every compress() call appends arch/FLOPs/latency/CPU info to ~/.dnaty/telemetry.jsonl. Seeds the GBM latency predictor. Never raises.

New Scripts

• scripts/build_latency_dataset.py — build GBM latency surrogate from measured ONNX timings.

Install

pip install dnaty==2.0.0

Benchmarks

Dataset	FLOPs ↓	Val acc
Dry Bean Quality	−83.4%	92.43%
Predictive Maint. (AI4I)	−83.1%	96.70%
Electrical Fault Detect	−83.0%	99.25%
Credit Fraud FULL (284K)	−64.0%	99.96%
HAR Sensors	−46.8%	99.17%
MNIST (full 70K)	−41.8%	98.68%

Continual learning (Split-MNIST, 5 tasks): 4.9× less forgetting than EWC (BWT −0.204 vs −0.998).

Full changelog

See CHANGELOG.md for complete details.